Deep learning based cricket batting shot classification and performance analysis using computer vision Authors: S. Rajarajeswari, Devika Prashant, Jayanth Srinivasan, Tejaswini Dubey Published: 2026 ABSTRACT Cricket is a sport which is played and known worldwide. It requires proper technique and execution of shots by the batsman. This research aims on classifying different types of batting shots and analysing the player’s performance using computer vision and deep learning techniques. It also includes comparison with the shots played by professional players. A detailed methodology explains about data collection, preprocessing, feature extraction, shot classification, and performance evaluation. Videos of players performing the batting shots is the source of data collection process. Optical flow calculation and pose estimation are the techniques used in the preprocessing phase to find the bat’s motion, the player’s position, and the shot’s trajectory. The features are given to a 3D Convolutional Neural Network which recognizes and classifies the different types of shots. The classified shot types are further used for shot comparison and analysis, helping us to provide a detailed analysis of the player’s performance. Moreover, the evaluation metrics and performance assessment are done by comparing with the shots played by top professionals that give the insights into the areas of improvement. The system is applicable in sports training, player performance analysis, and coaching, giving real-time insights and data-centric feedback for cricket players of all levels. Further developments could incorporate high-level biomechanics evaluation, real-time feedback implementation, and AI coaching support to allow for the further enhancement of batting performance. Introduction Cricket is a game that requires exceptional technical skill, good awareness of the situation, and excellent motor skills. The performance assessment in cricket includes various aspects, among which batting technique is still the most complicated and important to the experts and coaches. The conventional method of coaching depends on the manual video analysis of a batsman as the only way to evaluate his/her performance. But, thanks to the technological advancements in machine vision, deep learning, and pose estimation, the process of automating and objectively evaluating the players’ actions is growing more and more. The automatic classification of cricket shots and the analysis of player movements have turned into hot topics of research. A number of the latest investigations have focused on applying machine learning and deep learning methods over the video data to classify cricket shots. Suryavanshi et al.1 and Kumar et al.2 suggested a hybrid CNN-SVM and CNN-SoftMax approach for effective cricket shot classification, showing significant increases in recognition accuracy. Previous work by Karmaker et al.3 used motion vector-based analysis, which was an important step in the development of motion dynamics for sports video analytics. Recent studies have made use of deep learning models, especially convolutional neural networks (CNNs), to obtain spatial features in video frames4–6. The trained model of Khan et al.5 is based on deep neural networks which utilized static frame-level image data for the purpose of classifying different types of shots, whereas Foysal et al.6 proposed Shot-Net, a convolutional neural network that was built only for the recognition of various types of batting shots. Vision Transformers (ViT) have also been considered as possible solutions in this domain; Dey and Biswas7 released Shot-ViT, a model that harnesses self-attention methods to enhance classification accuracy even when dealing with longer time intervals. Optical flow methods have proven to be effective in capturing dynamic characteristics like the movement of the human body and the trajectory of a bat. Semwal et al.8 proposed applying optical flow to cricket for shot detection and performed motion segmentation. Likewise, the Lucas-Kanade and Farneback methods based on optical flow have been shown to model minor and major-scale motion, respectively, and thus, these methods are appropriate for tracking bat motion and player displacement. Pose estimation has increasingly become one of the key factors in the biomechanical analysis of cricket. The YOLOPose model enables keypoint detection at a very low latency which in turn allows the use of postures for tracking batting positions in a frame by frame manner. To illustrate, Siddiqui et al.9 and Khan et al.10 were able to explicitly point out the involvement of human pose in activity recognition and quality evaluation of the batting shots, respectively. Besides, the experimental setup suggested by Dias et al.11 and Peploe et al.12 has also indicated that the biomechanical data could be applied through and predict performance with machine learning models. Considering the improvements, the proposed system integrates multiple advanced technologies to classify cricket batting shots and measure professional players’ performance indicators. The system design incorporates a 3D Convolutional Neural Network for extracting spatial and temporal information from the video data, optical flow algorithms for determining the motion vectors, and YOLOPose for the estimation of the pose at the frame level. Moreover, the feature vectors obtained from the network are then compared using cosine similarity to determine how closely the batting techniques of the players are similar. The model improves the player comparison analysis and also helps coaches, sports analysts and athletes to assess their techniques. It also allows the players to find areas for improvement and making necessary strategic decisions. The overall research contributes in improving sports analytics to a higher level. Literature survey The progress made in the technologies of computer vision and deep learning fields have led to an amazing improvement in the classification of cricket shots and the analysis of players’ performances. Deep learning models for shot classification Convolutional neural networks and hybrid models are the most used models in research studies regarding batting shot classification. Suryavanshi et al.1 mentioned that a CNN-SVM hybrid model produced 97.16% recall and 91.08% accuracy. Kumar et al.2 compared the performance of two models, CNN-SoftMax and CNN-SVM. It was found that CNN-SVM was better than CNN-softmax with an accuracy of 91.37%. Karmaker et al.3 used motion vector features to classify batting shots. Bhat et al.4 explained the importance of transfer learning by using pretrained models such as VGG19 and ResNet for video based classification. In the context of real-time applicability, Semwal et al.8 implemented a CNN-based technique for cricket shot detection from videos of live matches. Khan et al.5 showed the practical applicability of frame based learning throughout full video sequences as they worked with static image data and deep neural networks. The custom CNN developed by Foysal et al.6, Shot-Net, which was trained on annotated cricket datasets, was able to successfully classify the shots. Vision transformers and temporal modeling Transformer networks are being investigated in recent works to improve context modeling. Shot-ViT, a Vision Transformer-based model developed on the Cricket Batting Shots Image Dataset (CBSId), was proposed by Dey and Biswas7. It outperformed CNN-based models in capturing global spatial dependencies, with an accuracy of 92.58%. Ahmad et al.13 proposed a hybrid model consisting of time distributed 2D CNN and LSTM units to handle temporal dependencies. The model was able to classify complex movements such as pull shots and reverse sweeps with an accuracy of 92%. Pose estimation and keypoint-based classification Pose estimation is one of the most dependable methods for checking the movement of people in sports. With an accuracy of 99.77%, Siddiqui et al.9 classified and distinguished the different types of cricket strokes through SVM, Random Forest and LSTM models while MediaPipe helped in the keypoints extraction of the body. This union of pose estimation and machine learning gives excellent results in tracking the performance of players and also in coaching. Biomechanics and perception-action coupling Biomechanical studies give important information about a player’s performance apart from video analysis. Using Vicon motion capture systems, Peploe et al.12 studied how batsmen reacted to different delivery techniques and discovered considerable kinematic distinctions. By using a temporal occlusion model, Mann et al.14 explored anticipatory performance, stressing the importance of perception action coupling in proficient batting. In order to bridge the gap between lab-based biomechanics and real-game situations, Dias et al.11 suggested experimental protocols that use high-speed motion capture to predict shot execution. Effective gaze control has a direct impact on shot timing and success, according to research by Land and McLeod15 on visual- motor coordination in batting. Wearable sensor-based recognition Sensor-driven methods provide useful substitutes for gathering data outside of a lab. In order to categorize 24 batting shots according to motion dynamics, Khan et al.10 created a hierarchical recognition system using IMUs. Their system’s F1-score of more than 88% validates its application in low-level training settings. Multi-modal and hybrid systems In order to classify batting shots using 800 shot clips, Khan et al.16 combined 2D CNNs with LSTMs and contrasted them with 3D CNNs. According to their findings, 3D CNNs perform better than sequential models because they are better at capturing spatial-temporal features. A CNN-GRU framework (Cricshotclassify) was proposed by Sen et al.17 to classify strokes from video sequences with state-of-the-art accuracy. Surveys and frameworks Cricket performance analysis using deep learning methods has been discussed in detail by Mahajan et al.18 and Akhilesh et al.19. They have pointed out the trends, unresolved issues and new methods present in it. One of the most significant aspects of these reviews is the transition from the use of manually crafted features to automated deep learning pipelines. In the end, Chauhan and Bhatia20 showed that computer vision methods can be applied to identify general cricket actions, which in turn creates the base for more specific tasks such as recognizing shots. Dataset description We made use of the publicly available “Cricket Shots-IPL 2023” dataset for our research, which was obtained from Kaggle (https://www.kaggle.com/datasets/taarunsridhar/cricket-shots-ipl-2023). The dataset consists of short video clips of different cricket batting shots from the Indian Premier League 2023 season. The dataset was curated and preprocessed so that it can be overcome class imbalance. The dataset was balanced such that each shot category had around 150 videos.This makes the dataset uniform. The classification is done based on the following shot categories: Slog, sweep, drive, pull, misc, cut, and flick These categories were chosen to ensure that the model can distinguish between shots by learning the fine differences, as it would be trained on a large variety of batting shots. Along with the primary dataset, we constructed a custom reference dataset that consisted of carefully curated videos of top professional cricket players executing the same shot types. These reference videos were collected through manual work from highlight reels and match footage that were made public. The reference dataset was used to calculate cosine similarity between the predicted shot of the player and those of the top players, which improved performance comparison, as well as shot improvement analysis, thus providing insights through the comparison and analysis of shot improvement. Methodology The methodology consists of various phases starting from data collection to the final analysis, as illustrated in Fig. 1. A comprehensive description of each phase in the methodology is explained below. Fig. 1System architecture. System architecture. Data collection The first step in the methodology is to collect different cricket batting videos and to group them with significant labels for supervised learning. Video acquisition The primary dataset-Cricket Shots IPL 2023-dataset is obtained from Kaggle. This dataset contains short video clips of professional batting shots extracted from IPL match footages. The intent was to cover a wide range of shot types like drive, pull, cut, slog, sweep, flick, and many others. Class balancing The dataset was made such that each of the categories including drive, pull, cut, sweep, flick, slog, and miscellaneous had a minimum of 150 videos. This step assured learning, balancing and class bias prevention. Such a distribution assists the model in acquiring uniform representations for every shot type. Annotation Every video clip was carefully classified into its respective shot category using visual inspection. These annotations were considered ground truths during the training and testing phases. Preprocessing The second step in the methodology is Data Preprocessing. The following preprocessing steps have been applied to the video samples collected: Frame extraction Using OpenCV, video samples in the dataset was transformed such that each video has a certain number of uniformly distributed frames. Frame extraction process is useful in helping the model learn about motion and positional information efficiently. In order to make the input videos consistent, the frames are normalized and resized to a common resolution. Optical flow calculation The Farneback Dense Optical Flow algorithm was implemented using OpenCV to capture motion dynamics between consecutive frames. The algorithm was configured with the following parameters to optimize motion detection for cricket batting analysis: Pyramid Scale: 0.5 (image scale for building pyramids).Pyramid Levels: 3 (number of pyramid layers).Window Size: 15 pixels (averaging window size).Iterations: 3 (number of iterations at each pyramid level).Polynomial Neighborhood: 5 (size of pixel neighborhood).Polynomial Sigma: 1.2 (standard deviation of Gaussian for derivative calculation).Flags: 0 (default Farneback flags). Pyramid Scale: 0.5 (image scale for building pyramids). Pyramid Scale: 0.5 (image scale for building pyramids). Pyramid Levels: 3 (number of pyramid layers). Pyramid Levels: 3 (number of pyramid layers). Window Size: 15 pixels (averaging window size). Window Size: 15 pixels (averaging window size). Iterations: 3 (number of iterations at each pyramid level). Iterations: 3 (number of iterations at each pyramid level). Polynomial Neighborhood: 5 (size of pixel neighborhood). Polynomial Neighborhood: 5 (size of pixel neighborhood). Polynomial Sigma: 1.2 (standard deviation of Gaussian for derivative calculation). Polynomial Sigma: 1.2 (standard deviation of Gaussian for derivative calculation). Flags: 0 (default Farneback flags). Flags: 0 (default Farneback flags). These parameters were selected to capture both fine-grained bat movements and larger body motions during shot execution. The optical flow computation yielded motion magnitude and directional angle for each pixel, which were aggregated to extract motion feature vectors representing bat speed, swing direction, and temporal motion patterns across the shot sequence. Pose estimation The YOLOv8-Pose model was used for the keypoint detection. It helps us to identify 17 major body joints from each frame, which includes the shoulders, elbows, wrists, knees, and ankles. It then provides a skeletal keypoints structure which helps the model to analyse the player’s posture better. It also helps the model to understand about the form of the shot from the features extracted from the structural data. These minor differences at the main keypoints of the body helps the model understand and predict the shot type with more clarity. This helps in distinguishing very similar types of shot more accurately. The YOLOv8-Pose model was configured as follows: Model Variant: YOLOv8n-pose (nano variant for computational efficiency).Inference Speed: 90.3ms per image (as demonstrated in Fig. 3).Keypoint Format: 17-keypoint COCO skeleton structure.Detected Keypoints: Shoulders, elbows, wrists, hips, knees, ankles, and other major body joints. Model Variant: YOLOv8n-pose (nano variant for computational efficiency). Model Variant: YOLOv8n-pose (nano variant for computational efficiency). Inference Speed: 90.3ms per image (as demonstrated in Fig. 3). Inference Speed: 90.3ms per image (as demonstrated in Fig. 3). Keypoint Format: 17-keypoint COCO skeleton structure. Keypoint Format: 17-keypoint COCO skeleton structure. Detected Keypoints: Shoulders, elbows, wrists, hips, knees, ankles, and other major body joints. Detected Keypoints: Shoulders, elbows, wrists, hips, knees, ankles, and other major body joints. The lightweight nano variant balances detection accuracy with processing speed, achieving real-time pose detection with confidence scores ranging from 0.57 to 0.86 across multiple players in broadcast footage, as shown in the experimental results. Feature extraction The features are extracted from video samples using Optical Flow and Pose Estimation techniques. These features are used for batting shot type classification and in evaluating the player’s performance. Feature fusion architecture:The multimodal feature fusion process combines spatial, temporal, and biomechanical information from three sources: Feature fusion architecture:The multimodal feature fusion process combines spatial, temporal, and biomechanical information from three sources: Feature fusion architecture: The multimodal feature fusion process combines spatial, temporal, and biomechanical information from three sources: Raw RGB frames:Input Shape: (20 frames, 75 × 75 pixels, 3 channels)Normalized to [0, 1] range.Directly fed to 3D CNN for spatiotemporal learning. Raw RGB frames:Input Shape: (20 frames, 75 × 75 pixels, 3 channels)Normalized to [0, 1] range.Directly fed to 3D CNN for spatiotemporal learning. Raw RGB frames:Input Shape: (20 frames, 75 × 75 pixels, 3 channels)Normalized to [0, 1] range.Directly fed to 3D CNN for spatiotemporal learning. Input Shape: (20 frames, 75 × 75 pixels, 3 channels) Input Shape: (20 frames, 75 × 75 pixels, 3 channels) Normalized to [0, 1] range. Normalized to [0, 1] range. Directly fed to 3D CNN for spatiotemporal learning. Directly fed to 3D CNN for spatiotemporal learning. (b)Optical flow features (motion-based):Extracted per frame: motion magnitude computed using Farneback algorithm.Aggregated motion intensity: mean motion magnitude across all pixels per frame.Temporal sequence: 20 motion intensity values representing shot dynamics.Feature vector: 20-dimensional motion profile capturing bat swing timing. (b)Optical flow features (motion-based):Extracted per frame: motion magnitude computed using Farneback algorithm.Aggregated motion intensity: mean motion magnitude across all pixels per frame.Temporal sequence: 20 motion intensity values representing shot dynamics.Feature vector: 20-dimensional motion profile capturing bat swing timing. Optical flow features (motion-based):Extracted per frame: motion magnitude computed using Farneback algorithm.Aggregated motion intensity: mean motion magnitude across all pixels per frame.Temporal sequence: 20 motion intensity values representing shot dynamics.Feature vector: 20-dimensional motion profile capturing bat swing timing. Extracted per frame: motion magnitude computed using Farneback algorithm. Extracted per frame: motion magnitude computed using Farneback algorithm. Aggregated motion intensity: mean motion magnitude across all pixels per frame. Aggregated motion intensity: mean motion magnitude across all pixels per frame. Temporal sequence: 20 motion intensity values representing shot dynamics. Temporal sequence: 20 motion intensity values representing shot dynamics. Feature vector: 20-dimensional motion profile capturing bat swing timing. Feature vector: 20-dimensional motion profile capturing bat swing timing. (c)Pose keypoint features (biomechanical):Per frame: 17 keypoints × 2 coordinates (x, y) = 34 values.Temporal concatenation: 20 frames × 34 = 680 features.Normalized to [0, 1] range based on frame dimensions.Feature vector: 680-dimensional pose trajectory. (c)Pose keypoint features (biomechanical):Per frame: 17 keypoints × 2 coordinates (x, y) = 34 values.Temporal concatenation: 20 frames × 34 = 680 features.Normalized to [0, 1] range based on frame dimensions.Feature vector: 680-dimensional pose trajectory. Pose keypoint features (biomechanical):Per frame: 17 keypoints × 2 coordinates (x, y) = 34 values.Temporal concatenation: 20 frames × 34 = 680 features.Normalized to [0, 1] range based on frame dimensions.Feature vector: 680-dimensional pose trajectory. Per frame: 17 keypoints × 2 coordinates (x, y) = 34 values. Per frame: 17 keypoints × 2 coordinates (x, y) = 34 values. Temporal concatenation: 20 frames × 34 = 680 features. Temporal concatenation: 20 frames × 34 = 680 features. Normalized to [0, 1] range based on frame dimensions. Normalized to [0, 1] range based on frame dimensions. Feature vector: 680-dimensional pose trajectory. Feature vector: 680-dimensional pose trajectory. (d)Fusion strategy:The system employs implicit fusion where optical flow and pose features guide the analysis pipeline:RGB frames are processed through 3D CNN for primary classification.Optical flow features are used for motion-based shot comparison (cosine similarity).Pose features enhance interpretability and provide biomechanical context.This multi-stream approach enables both classification and performance evaluation. (d)Fusion strategy:The system employs implicit fusion where optical flow and pose features guide the analysis pipeline:RGB frames are processed through 3D CNN for primary classification.Optical flow features are used for motion-based shot comparison (cosine similarity).Pose features enhance interpretability and provide biomechanical context.This multi-stream approach enables both classification and performance evaluation. Fusion strategy: The system employs implicit fusion where optical flow and pose features guide the analysis pipeline:RGB frames are processed through 3D CNN for primary classification.Optical flow features are used for motion-based shot comparison (cosine similarity).Pose features enhance interpretability and provide biomechanical context.This multi-stream approach enables both classification and performance evaluation. RGB frames are processed through 3D CNN for primary classification. RGB frames are processed through 3D CNN for primary classification. Optical flow features are used for motion-based shot comparison (cosine similarity). Optical flow features are used for motion-based shot comparison (cosine similarity). Pose features enhance interpretability and provide biomechanical context. Pose features enhance interpretability and provide biomechanical context. This multi-stream approach enables both classification and performance evaluation. This multi-stream approach enables both classification and performance evaluation. (2)Normalization: (2)Normalization: Normalization: The feature values obtained from optical flow and Pose estimation techniques are normalized. This helps in eliminating any external noise, lighting variations and differences in camera angles. Now the model easily focuses on the important features alone. (3)Feature normalization: (3)Feature normalization: Feature normalization: All feature modalities were normalized to ensure uniform scale: RGB frames: Pixel values normalized to [0, 1] by dividing by 255.Optical flow magnitudes: Inherently scaled by motion intensity.Pose keypoints: Coordinates normalized by frame dimensions. RGB frames: Pixel values normalized to [0, 1] by dividing by 255. RGB frames: Pixel values normalized to [0, 1] by dividing by 255. Optical flow magnitudes: Inherently scaled by motion intensity. Optical flow magnitudes: Inherently scaled by motion intensity. Pose keypoints: Coordinates normalized by frame dimensions. Pose keypoints: Coordinates normalized by frame dimensions. This normalization prevents any single feature modality from dominating the learning process and ensures stable gradient flow during backpropagation. Model development In order to classify cricket shots based on the temporal and spatial patterns found in the frame sequences, a deep learning model was developed using features extracted from the video data. 3D CNN architecture The video data was treated as 3D volumetric data during the processing, as a 3D Convolutional Neural Network was applied. This model is excellent for player movement analysis, as it captures both time and space dimensions, which is a limitation of 2D CNNs that work only in the spatial domain. The architecture comprises of: To gradually learn more abstract features, several 3D convolutional layers with progressively larger filter sizes (32 → 64 → 128 → 256) are used.After every convolution, batch normalization is performed to speed up and stabilize learning.3D max pooling layers that preserved temporal context while reducing spatial dimensions.The 3D feature maps were compressed into a compact representation using a Global Average Pooling layer.Two dense layers (512 and 256 units) that are fully connected and have dropout for regularization.A final softmax layer for outputting shot class probabilities. To gradually learn more abstract features, several 3D convolutional layers with progressively larger filter sizes (32 → 64 → 128 → 256) are used. To gradually learn more abstract features, several 3D convolutional layers with progressively larger filter sizes (32 → 64 → 128 → 256) are used. After every convolution, batch normalization is performed to speed up and stabilize learning. After every convolution, batch normalization is performed to speed up and stabilize learning. 3D max pooling layers that preserved temporal context while reducing spatial dimensions. 3D max pooling layers that preserved temporal context while reducing spatial dimensions. The 3D feature maps were compressed into a compact representation using a Global Average Pooling layer. The 3D feature maps were compressed into a compact representation using a Global Average Pooling layer. Two dense layers (512 and 256 units) that are fully connected and have dropout for regularization. Two dense layers (512 and 256 units) that are fully connected and have dropout for regularization. A final softmax layer for outputting shot class probabilities. A final softmax layer for outputting shot class probabilities. Training process Dataset partitioning: The Cricket Shots-IPL 2023 dataset containing approximately 1050 videos (150 per shot class across 7 categories) was split as follows: Training Set: 1021 video samples (approximately 80%).Test Set: 752 video samples (approximately 20%).Validation: Test set used for validation during training Input Specifications:Frames per Video: 20 (uniformly sampled across video duration).Frame Resolution: 75 × 75 pixels .Color Channels: 3 (RGB).Input Tensor Shape: (20, 75, 75, 3) Optimizer and Hyperparameters:Optimizer: AdamW (Adam with weight decay regularization).Initial Learning Rate: 0.001.Weight Decay: 0.0001.Batch Size: 16.Total Epochs: 50.Loss Function: Categorical Cross-Entropy.Activation Functions: ReLU (convolutional and dense layers), Softmax (output layer). Training Set: 1021 video samples (approximately 80%). Training Set: 1021 video samples (approximately 80%). Test Set: 752 video samples (approximately 20%). Test Set: 752 video samples (approximately 20%). Validation: Test set used for validation during training Input Specifications: Validation: Test set used for validation during training Input Specifications: Frames per Video: 20 (uniformly sampled across video duration). Frames per Video: 20 (uniformly sampled across video duration). Frame Resolution: 75 × 75 pixels . Frame Resolution: 75 × 75 pixels . Color Channels: 3 (RGB). Color Channels: 3 (RGB). Input Tensor Shape: (20, 75, 75, 3) Optimizer and Hyperparameters: Input Tensor Shape: (20, 75, 75, 3) Optimizer and Hyperparameters: Optimizer: AdamW (Adam with weight decay regularization). Optimizer: AdamW (Adam with weight decay regularization). Initial Learning Rate: 0.001. Initial Learning Rate: 0.001. Weight Decay: 0.0001. Weight Decay: 0.0001. Batch Size: 16. Batch Size: 16. Total Epochs: 50. Total Epochs: 50. Loss Function: Categorical Cross-Entropy. Loss Function: Categorical Cross-Entropy. Activation Functions: ReLU (convolutional and dense layers), Softmax (output layer). Activation Functions: ReLU (convolutional and dense layers), Softmax (output layer). 3D CNN architecture details: Convolutional filter progression: 32 → 64 → 128 → 256.Kernel size: (3, 3, 3) for all Conv3D layers.Pooling strategy:Layers 1–2: (1, 2, 2) - preserves temporal dimension while reducing spatial.Layers 3–4: (2, 2, 2) - reduces both temporal and spatial dimensions.Batch normalization: Applied after each convolutional layer.Global average pooling: Reduces 3D feature maps to 1D vector.Fully connected layers: 512 → 256 → 7 neurons.Dropout: 0.4 applied after dense layers for regularization.Total parameters: 1,430,919 (1,429,959 trainable, 960 non-trainable). Convolutional filter progression: 32 → 64 → 128 → 256. Convolutional filter progression: 32 → 64 → 128 → 256. Kernel size: (3, 3, 3) for all Conv3D layers. Kernel size: (3, 3, 3) for all Conv3D layers. Pooling strategy:Layers 1–2: (1, 2, 2) - preserves temporal dimension while reducing spatial.Layers 3–4: (2, 2, 2) - reduces both temporal and spatial dimensions. Pooling strategy:Layers 1–2: (1, 2, 2) - preserves temporal dimension while reducing spatial.Layers 3–4: (2, 2, 2) - reduces both temporal and spatial dimensions. Layers 1–2: (1, 2, 2) - preserves temporal dimension while reducing spatial. Layers 1–2: (1, 2, 2) - preserves temporal dimension while reducing spatial. Layers 3–4: (2, 2, 2) - reduces both temporal and spatial dimensions. Layers 3–4: (2, 2, 2) - reduces both temporal and spatial dimensions. Batch normalization: Applied after each convolutional layer. Batch normalization: Applied after each convolutional layer. Global average pooling: Reduces 3D feature maps to 1D vector. Global average pooling: Reduces 3D feature maps to 1D vector. Fully connected layers: 512 → 256 → 7 neurons. Fully connected layers: 512 → 256 → 7 neurons. Dropout: 0.4 applied after dense layers for regularization. Dropout: 0.4 applied after dense layers for regularization. Total parameters: 1,430,919 (1,429,959 trainable, 960 non-trainable). Total parameters: 1,430,919 (1,429,959 trainable, 960 non-trainable). Training callbacks: ReduceLROnPlateau:Monitors: validation loss.Reduction factor: 0.5 (halves learning rate).Patience: 5 epochs.Applied at epochs 6, 20, and 33.Final learning rate: 0.000125.EarlyStopping:Monitors: validation loss.Patience: 10 epochs.Restores best model weights automatically. ReduceLROnPlateau:Monitors: validation loss.Reduction factor: 0.5 (halves learning rate).Patience: 5 epochs.Applied at epochs 6, 20, and 33.Final learning rate: 0.000125. ReduceLROnPlateau:Monitors: validation loss.Reduction factor: 0.5 (halves learning rate).Patience: 5 epochs.Applied at epochs 6, 20, and 33.Final learning rate: 0.000125. Monitors: validation loss. Monitors: validation loss. Reduction factor: 0.5 (halves learning rate). Reduction factor: 0.5 (halves learning rate). Patience: 5 epochs. Patience: 5 epochs. Applied at epochs 6, 20, and 33. Applied at epochs 6, 20, and 33. Final learning rate: 0.000125. Final learning rate: 0.000125. EarlyStopping:Monitors: validation loss.Patience: 10 epochs.Restores best model weights automatically. EarlyStopping:Monitors: validation loss.Patience: 10 epochs.Restores best model weights automatically. Monitors: validation loss. Monitors: validation loss. Patience: 10 epochs. Patience: 10 epochs. Restores best model weights automatically. Restores best model weights automatically. Training completed successfully after 50 epochs with final training accuracy of 97.03% and validation accuracy of 95.61%. Data augmentation Training videos were improved using: Horizontal flipping.Brightness variation.Cropping. Horizontal flipping. Horizontal flipping. Brightness variation. Brightness variation. Cropping. Cropping. The model was able to generalize to new, unseen video clips more effectively through these steps. Classification The unseen cricket shot videos were classified through the 3D CNN model after the training. Shot prediction The model outputted a probability distribution over the set shot classes: drive, pull, cut, sweep, flick, slog, and miscellaneous, for each test video. The predicted label was taken from the class with the highest probability. Performance evaluation The performance of the classification model was evaluated using the following: Accuracy - It tells the number of correctly predicted test videos among all the video samples.Precision and Recall - It shows the model’s performance in differentiating the various batting shot types.F1-Score - Found by calculating the harmonic mean between precision and recall. It also helps in balancing the number of false positives and false negatives.Confusion Matrix - Matrix that explains the frequency of a particular shot type being misclassified as another shot type. Accuracy - It tells the number of correctly predicted test videos among all the video samples. Accuracy - It tells the number of correctly predicted test videos among all the video samples. Precision and Recall - It shows the model’s performance in differentiating the various batting shot types. Precision and Recall - It shows the model’s performance in differentiating the various batting shot types. F1-Score - Found by calculating the harmonic mean between precision and recall. It also helps in balancing the number of false positives and false negatives. F1-Score - Found by calculating the harmonic mean between precision and recall. It also helps in balancing the number of false positives and false negatives. Confusion Matrix - Matrix that explains the frequency of a particular shot type being misclassified as another shot type. Confusion Matrix - Matrix that explains the frequency of a particular shot type being misclassified as another shot type. Comparison of metrics Calculation of motion-based metrics Optical flow technique gives motion feature vectors. This was used to track the bat movements through each frames. These vectors help us analyse how a particular shot becomes better or worse by finding the motion’s speed and direction at each frame. Cosine similarity calculation Cosine similarity helps us know the similarity between the player’s shot and the shot played by a professional reference player. It always gives a score between 0 and 1 which tells how similar their shot timing and motion are. Higher scores indicate more accurate body motion and bat swing rhythm alignment. Interpretive feedback The following were the feedbacks for different cases respectively: “Excellent shot timing!” was generated for scores higher than 0.85.“Decent timing, improve consistency” was indicated by scores ranging from 0.75 to 0.85.“Poor bat speed or delayed execution” indicated by scores below 0.6 to 0.75. “Excellent shot timing!” was generated for scores higher than 0.85. “Excellent shot timing!” was generated for scores higher than 0.85. “Decent timing, improve consistency” was indicated by scores ranging from 0.75 to 0.85. “Decent timing, improve consistency” was indicated by scores ranging from 0.75 to 0.85. “Poor bat speed or delayed execution” indicated by scores below 0.6 to 0.75. “Poor bat speed or delayed execution” indicated by scores below 0.6 to 0.75. Output and analysis The last part showcases the outputs and its analysis. Visualization Various visual outputs were created for better understanding of the model performance and actions of individual shots: Confusion Matrix indicating the correctness of the classification for all shot types.Accuracy Curves for comparing the progress of training and validation.Body Pose Visualizations by YOLO for body posture frame by frame.Optical Flow Maps depicting motion dynamics during the various stages of a shot (e.g., stance, swing, follow-through).Screenshots of Cosine Similarity Output for personalized feedback. Confusion Matrix indicating the correctness of the classification for all shot types. Confusion Matrix indicating the correctness of the classification for all shot types. Accuracy Curves for comparing the progress of training and validation. Accuracy Curves for comparing the progress of training and validation. Body Pose Visualizations by YOLO for body posture frame by frame. Body Pose Visualizations by YOLO for body posture frame by frame. Optical Flow Maps depicting motion dynamics during the various stages of a shot (e.g., stance, swing, follow-through). Optical Flow Maps depicting motion dynamics during the various stages of a shot (e.g., stance, swing, follow-through). Screenshots of Cosine Similarity Output for personalized feedback. Screenshots of Cosine Similarity Output for personalized feedback. Feedback messages The following are the feedback given by the model based on similarity scores respectively: “Excellent shot timing!” for high similarity.“Improve reaction speed” for moderate scores.“Focus on bat speed and follow-through” for low scores. “Excellent shot timing!” for high similarity. “Excellent shot timing!” for high similarity. “Improve reaction speed” for moderate scores. “Improve reaction speed” for moderate scores. “Focus on bat speed and follow-through” for low scores. “Focus on bat speed and follow-through” for low scores. The feedback messages served as a link between the numerical analysis and user interpretation, thus making the system available not only to players but also to coaches. Interpretation and insight The classification and comparison outputs, when merged together, allowed for an all-encompassing evaluation of a batsman’s performance. Not only they would know what shot they played, but also how good they were at doing it when compared to the best players. This way, the system could be used for live coaching, post-match analysis, and planning of training sessions. Research gap The cricket shot analysis field has been reliant on the manual watching of games and expert coaching for the longest time, which resulted in subjective and inconsistent evaluations. On the other hand, technologies such as Hawk- Eye and motion capturing systems offer comprehensive analytics but often are pricey and out of reach for amateur players. Current works in sports analytics and computer vision have mainly dealt with player tracking, action recognition, and biomechanical analysis, but there is no existence of an AI-powered automated shot analyzing system that gives detailed and objectified comparisons of professional and amateur cricketers. In addition, there have been studies on deep learning models like 3D CNNs, LSTMs, and Transformers in general action recognition, but their use in cricket-specific batting shot analysis is still very limited. The same can be said for the accuracy of metrics such as bat speed, bat angle, and shot execution quality, which are derived through the combination of pose estimation, optical flow, and deep learning models. Overcoming this barrier can make cricket training really effective and monitor the performance of top players in such a way that it accessible to all players. Result The proposed cricket batting shot classification system achieved a test accuracy of 95.61% on 752 unseen video samples from the Cricket Shots-IPL 2023 dataset. The model demonstrated consistent performance across all seven shot categories (pull, flick, misc, cut, slog, sweep, drive) with macro-averaged precision, recall, and F1-scores all at 0.95. Training converged after 50 epochs with a final training accuracy of 95.49%, indicating excellent generalization without overfitting. The following subsections detail the optical flow analysis, pose detection validation, model architecture specifications, and per-class classification performance. Model performance and evaluation metrics In order to assess the time evolution of cricket shots, the Farneback algorithm was applied to compute optical flow over three representative video frames (Frame 0, 10, and 20). The results illustrated in Fig. 2a–c indicate that Frame 0 shows the batsman’s initial position at a relatively low average motion magnitude of 1.18 and at a direction angle of 120.98°, which suggests very slight body and bat movement during the shot preparation. In Frame 10, the motion intensity (magnitude: 1.85) has a big increase and the directional angle has a shift to 155.25°, which marks the main execution phase where the bat is vigorously swung to hit the ball. At the end of the motion (by Frame 20), the direction is the opposite (angle: 192.02°) and though the motion magnitude is still quite large (1.74) the pattern is that of a gradual decrease typical of the follow-through phase. The optical flow distinguish the motion stages required for the shot type classification and performance assessment of players through the change in the flow of time in both magnitude and direction. Fig. 2Optical flow analysis (a) Frame 0 – Pre-shot stance, (b) Frame 10 – Shot execution, (c) Frame 20 – Follow-through phase. Optical flow analysis (a) Frame 0 – Pre-shot stance, (b) Frame 10 – Shot execution, (c) Frame 20 – Follow-through phase. The three optical flow frames show significant differences in batsman’s movement, posture, and body mechanics. The color-coded flow maps show the directional gradients and intensity of movement accurately which is necessary for extracting features on a frame basis. Motion capture then goes to the 3D CNN model directly where the model detects the patterns that are particular to different types of shots. Being able to perform thorough motion capture makes it easy to compare players by using cosine similarity. It helps in the spotting of timing mismatches, slow follow-throughs, or even differences in technique when compared to elite players. Consequently, this leads to the uncovering of the model’s capability to classify and to provide performance analytics that are feedback-based. Fig. 3Cricket player pose detection using YOLO-Pose. Cricket player pose detection using YOLO-Pose. Figure 3 depicts successful real-time pose detection via YOLOv8-Pose on a live frame from an IPL match broadcast. Four players have been identified with confidence scores that vary from 0.57 to 0.86, and the main joints, namely, shoulders, elbows, knees, and ankles, are well recognized. The keypoint annotations make it very clear that the batsman is in the position and the stance from which he is ready to hit, which can be further processed for the extraction of the biomechanical features of bat angle, knee bend, and body posture. This step is crucial in the case of differentiating the types of shots and judging the player’s skill, hence, allowing a thorough analysis of a player’s batting form to be done on a frame-by-frame basis. The quick inference time of 90.3 ms also indicates that the model can be used in real-time sports analytics systems as it is suitable for real-time operation. Fig. 43D CNN model architecture and layer-wise parameter summary. 3D CNN model architecture and layer-wise parameter summary. The suggested deep learning architecture is built around a 3D Convolutional Neural Network (3D CNN) that can efficiently recognize and analyze both spatial and temporal patterns in cricket shot video sequences. In Fig. 4, the model can be seen comprising four 3D convolutional layers which have (32 → 64 → 128 → 256) filter depths corresponding to the increasing number of filters, each being subsequently treated with batch normalization and max pooling operations in order to build stability in learning and reduce the chance of overfitting. The size of the space is reduced gradually with time context still being kept. Then forwarding to global average pooling layer which is followed by fully connected layers of sizes 512 and 256, annotated with dropout layers as a way of preventing overfitting. The last dense layer applies softmax activation to assign an input to one of the seven shot categories. The model is reported to have a total of 1,430,919 parameters, out of which 1,429,959 are trainable parameters, which is an indication of a very expressive yet compact architecture that is optimized for video-based sports classification. Fig. 5Training vs. validation accuracy curve of 3D CNN shot classifier. Training vs. validation accuracy curve of 3D CNN shot classifier. The accuracy curve in Fig. 5 illustrates the learning behavior of the 3D CNN shot classification model during the 50 epochs. The training accuracy steadily rises, reaching over 90% by the last epoch, thus confirming the effective learning from the training data. The validation accuracy, although initially fluctuating, stabilizes and significantly improves after epoch 30, particularly towards the end where it closely follows the training curve. Therefore, it can be inferred that the model has learned the features that are generalizable, thus resulting in less overfitting by employing techniques such as augmentation, dropout, and early stopping. The overlapping of training and validation accuracy corroborates the strength of the model for the real-world cricket shot classification tasks. Fig. 6Confusion matrix. Confusion matrix. The cricket shot classification model’s confusion matrix produced, as seen in Fig. 6, has proved the model’s exceptionally high-performance prediction of batting shots for various types very clearly. The accurate categories are indicated by the diagonal of the matrix, which shows that the model’s predicted labels are the same as the true ones since there are a lot of values along the diagonal. The model’s performance is so good that it can tell apart the distinctive cricketing strokes very well. It can accurately predict nearly all the key shot types: pull (99), flick (102), misc (108), cut (106), slog (110), sweep (83), and drive (109). The errors of the model are very few and dispersed; for instance, some pull shots were said to be cuts (6) or misc (2), fours were sometimes predicted as flicks (4), misc as drives (4), and sweeps as misc (1) or cuts (1). All this indicates the existence of certain areas where improvements can be made but they would be minor ones nonetheless. In addition, it is worth mentioning the dataset has pretty even distribution among the classes, the major categories contain mostly about 100 and even more samples, thus, it is a strong point for the evaluated metrics’ reliability. Performance like this hints that the model will be an exceptional one considering its ability to classify shots accurately and compare them with the elite players. Table 1Classification report.MetricPrecisionRecallF1-ScoreSupportPull0.960.910.93109Flick0.930.980.95104Misc0.990.920.96117Cut0.910.950.93111Slog0.960.990.97111Sweep0.980.980.9885Drive0.960.950.95115Accuracy––0.95752Macro avg0.950.950.95752Weighted avg0.950.950.95752 Table 1Classification report.MetricPrecisionRecallF1-ScoreSupportPull0.960.910.93109Flick0.930.980.95104Misc0.990.920.96117Cut0.910.950.93111Slog0.960.990.97111Sweep0.980.980.9885Drive0.960.950.95115Accuracy––0.95752Macro avg0.950.950.95752Weighted avg0.950.950.95752 Classification report. The classification report presented in Table 1 shows a strong and reliable performance in all shot categories, providing an overall accuracy of 95%. The main metrics that are precision, recall, and F1-score are all equally high, with both macro and weighted averages at 0.95, indicating that the model has achieved effectiveness that is not dependent on the class distribution. The model has especially high precision for “misc” (0.99) and “sweep” (0.98), and very good recall for “flick,” “slog,” and “sweep” (≥ 0.98), showing that there are very few false positives and false negatives for these classes. The “cut” shot has the lowest precision of 0.91, while “pull” has the lowest recall of 0.91, yet these are still considered as strong performance levels. The consistent performance across the evaluation criteria is a clear indication of the model’s reliability in correctly distinguishing between different cricket batting shots and thus, it can be confidently applied to sports analytics areas such as comparing players, analyzing shot patterns, and evaluating performance. Shot timing comparison and performance feedback Cosine similarity was used to motion intensity vectors obtained from optical flow in order to quantitatively compare a player’s batting shot with that of a professional. The motion features were extracted from both reference and target videos on the 20 frames that were evenly spaced in time. These vectors illustrate the timing of the shot, which includes the bat swing speed and the body movement rhythm. The cosine similarity score shows the similarity between the player’s shot timing and the reference shot. The similarity score obtained from the model after applying cosine similarity was found to be 86.8%. From this we can infer that the player’s shot and the reference shot was a close match. The model gives easily interpretable feedback, showing the player’s timing and any improvements required. Fig. 7Cosine similarity and feedback. Cosine similarity and feedback. The similarity score of 0.8679 shown in Fig. 7 is over the threshold 0.85, so the model labels the timing as “Excellent shot timing!“. This indicates that the player’s bat swing speed, initiation moment, and overall shot rhythm are very similar to those of professional players. It helps the player to assess their performance and encourages them to practice and get better. Comparative performance analysis To evaluate the effectiveness of our proposed multimodal 3D CNN approach, we contextualize our results against recent state-of-the-art methods in cricket shot classification literature, as summarized in Table 2. Table 2Performance comparison.MethodApproachAccuracyKey featuresReferenceCNN-SVM (Suryavanshi)Hybrid CNN-SVM91.08%Spatial features + SVM classifier 1 CNN-SVM (Kumar)Hybrid CNN-SVM91.37%CNN feature extraction + SVM 2 2D CNN + LSTM (Ahmad)Temporal modeling92.00%Time-distributed CNN + LSTM 13 Shot-ViT (Dey & Biswas)Vision Transformer92.58%Self-attention, global dependencies 7 Pose-based Classification (Siddiqui)Pose + ML classifiers99.77%MediaPipe keypoints + SVM/RF/LSTM 9 Proposed Method 3D CNN + Optical Flow + Pose 95.61 Spatiotemporal + Motion + Biomechanics This work Table 2Performance comparison.MethodApproachAccuracyKey featuresReferenceCNN-SVM (Suryavanshi)Hybrid CNN-SVM91.08%Spatial features + SVM classifier 1 CNN-SVM (Kumar)Hybrid CNN-SVM91.37%CNN feature extraction + SVM 2 2D CNN + LSTM (Ahmad)Temporal modeling92.00%Time-distributed CNN + LSTM 13 Shot-ViT (Dey & Biswas)Vision Transformer92.58%Self-attention, global dependencies 7 Pose-based Classification (Siddiqui)Pose + ML classifiers99.77%MediaPipe keypoints + SVM/RF/LSTM 9 Proposed Method 3D CNN + Optical Flow + Pose 95.61 Spatiotemporal + Motion + Biomechanics This work Performance comparison. Performance analysis Our multimodal approach demonstrates superior performance compared to existing single-modality and sequential processing methods: Superiority over hybrid CNN-SVM architectures Compared to CNN-SVM models achieving 91.08–91.37%1,2, our approach demonstrates a substantial 4.24–4.53% performance advantage. This improvement stems from the end-to-end learning capability of our 3D CNN architecture, which directly captures spatiotemporal patterns from video sequences rather than relying on two-stage feature extraction and classification. Traditional CNN-SVM approaches process frames independently, thereby losing the temporal continuity essential for distinguishing dynamic batting shots that differ primarily in execution timing and motion trajectory. Advantage over sequential temporal models The 2D CNN+LSTM approach by Ahmad et al13. achieves 92% accuracy through sequential frame processing with recurrent units. Our method demonstrates + 3.61% improvement, validating the effectiveness of 3D convolutional processing over sequential LSTM-based temporal modeling. The key distinction lies in 3D CNNs’ ability to learn spatiotemporal features through volumetric convolutions that simultaneously process spatial and temporal dimensions, capturing motion patterns more effectively than recurrent architectures that process temporal information sequentially. Furthermore, our explicit optical flow integration provides motion-specific representations that enhance discriminative power beyond what RGB frames alone can offer. Competitive with attention-based architectures Shot-ViT7, leveraging Vision Transformer architecture, achieves 92.58% accuracy through self-attention mechanisms that capture global spatial dependencies. Our 3D CNN-based approach outperforms this by + 3.03%, demonstrating that the inductive biases of convolutional architectures, specifically local connectivity and hierarchical feature learning, are particularly well-suited for video-based cricket shot classification. The combination of 3D convolutions with motion-specific optical flow features and biomechanical pose information provides a more comprehensive representation than attention-based spatial modeling alone. Contextualizing pose-based approaches While Siddiqui et al9. report 99.77% accuracy using pose-based classification with MediaPipe keypoints, this represents a specialized approach focused exclusively on biomechanical features extracted from controlled datasets. Our multimodal system balances multiple information sources, visual appearance, motion dynamics, and biomechanics, to achieve robust performance across diverse real-world match footage from IPL broadcasts, where lighting variations, camera angles, and occlusions present additional challenges. Architectural advantages of multimodal integration The superior performance of our approach is attributable to the synergistic integration of three complementary feature modalities: Spatiotemporal feature learning (3D CNN) The 3D convolutional architecture processes video as volumetric data with dimensions (time × height × width × channels), enabling the network to learn hierarchical spatiotemporal features. Our progressive pooling strategy (1 × 2 × 2 for early layers, 2 × 2 × 2 for deeper layers) preserves temporal information during initial feature extraction while gradually abstracting spatiotemporal patterns. This design captures both the visual appearance of shots and their temporal evolution, critical for distinguishing shots with similar stances but different execution dynamics. Motion-specific representations (optical flow) Farneback optical flow computation provides explicit motion magnitude and directional information, which serves two purposes: (1) enhancing the discriminative power of the classification model by providing motion cues complementary to RGB appearance, and (2) enabling quantitative shot timing comparison through cosine similarity with professional reference shots. This motion-based comparison framework offers objective performance metrics valuable for coaching applications. Biomechanical context (pose estimation) YOLOv8-Pose integration adds interpretable biomechanical features by tracking 17 body keypoints across frames. This provides skeletal trajectory information that is particularly discriminative for technique-dependent shots (sweeps, drives) where body positioning and joint angles are characteristic. The 90.3ms inference time ensures real-time applicability for live analysis scenarios. Per-class performance validation Examining the classification report (Table 1) and confusion matrix (Fig. 6) reveals the robustness of our multimodal approach: High-Motion Shots (Pull: 0.93 F1, Slog: 0.97 F1): These aggressive shots with rapid bat movement benefit from optical flow features that capture swing velocity and trajectory, resulting in excellent classification performance.Technique-Dependent Shots (Sweep: 0.98 F1, Drive: 0.95 F1): These shots requiring specific body positioning achieve the highest metrics, indicating that pose estimation features successfully capture characteristic biomechanical patterns.Versatile Shots (Flick: 0.95 F1, Cut: 0.93 F1): Consistently strong performance across diverse shot types validates the generalizability of the learned multimodal representations.Miscellaneous Category (0.96 F1): High precision (0.99) with slightly lower recall (0.92) reflects the model’s confidence in identifying this heterogeneous category while maintaining conservative classification to avoid false positives. High-Motion Shots (Pull: 0.93 F1, Slog: 0.97 F1): These aggressive shots with rapid bat movement benefit from optical flow features that capture swing velocity and trajectory, resulting in excellent classification performance. High-Motion Shots (Pull: 0.93 F1, Slog: 0.97 F1): These aggressive shots with rapid bat movement benefit from optical flow features that capture swing velocity and trajectory, resulting in excellent classification performance. Technique-Dependent Shots (Sweep: 0.98 F1, Drive: 0.95 F1): These shots requiring specific body positioning achieve the highest metrics, indicating that pose estimation features successfully capture characteristic biomechanical patterns. Technique-Dependent Shots (Sweep: 0.98 F1, Drive: 0.95 F1): These shots requiring specific body positioning achieve the highest metrics, indicating that pose estimation features successfully capture characteristic biomechanical patterns. Versatile Shots (Flick: 0.95 F1, Cut: 0.93 F1): Consistently strong performance across diverse shot types validates the generalizability of the learned multimodal representations. Versatile Shots (Flick: 0.95 F1, Cut: 0.93 F1): Consistently strong performance across diverse shot types validates the generalizability of the learned multimodal representations. Miscellaneous Category (0.96 F1): High precision (0.99) with slightly lower recall (0.92) reflects the model’s confidence in identifying this heterogeneous category while maintaining conservative classification to avoid false positives. Miscellaneous Category (0.96 F1): High precision (0.99) with slightly lower recall (0.92) reflects the model’s confidence in identifying this heterogeneous category while maintaining conservative classification to avoid false positives. The narrow F1-score range (0.93–0.98) across all seven categories demonstrates balanced performance without bias toward specific shot types, confirming effective multimodal feature integration. Discussion The proposed multimodal cricket batting shot classification system achieves a test accuracy of 95.61% with consistent precision, recall, and F1-scores of 0.95 across seven shot categories, representing a significant advancement in automated cricket analytics. This performance demonstrates the effectiveness of integrating spatiotemporal learning (3D CNNs), motion-specific features (optical flow), and biomechanical representations (pose estimation) for comprehensive shot recognition. Comparative positioning and performance advantages Our approach establishes new performance benchmarks compared to established methods in cricket shot classification: The 4–5% improvement over hybrid CNN-SVM architectures12 validates the effectiveness of end-to-end spatiotemporal learning versus two-stage feature extraction approaches. Traditional CNN-SVM methods extract spatial features from individual frames and rely on separate SVM classifiers, creating an optimization bottleneck that prevents joint learning of feature representations and classification boundaries. Our 3D CNN architecture circumvents this limitation through unified backpropagation across both feature extraction and classification stages. The 3.6% advantage over sequential 2D CNN+LSTM models13 confirms that volumetric 3D convolutions provide superior spatiotemporal modeling compared to recurrent architectures. While LSTMs process temporal information sequentially, introducing latency and potential vanishing gradient issues across long sequences, 3D CNNs apply convolutional filters across spatiotemporal volumes, enabling parallel processing and direct gradient flow across temporal dimensions. This architectural choice proves particularly effective for cricket shots, where key discriminative information often appears in brief temporal windows (e.g., bat-ball contact moment). The 3% improvement over Vision Transformer architectures7 demonstrates that convolutional inductive biases, local connectivity, translation equivariance, and hierarchical feature learning, remain highly effective for video-based action recognition despite the recent success of attention mechanisms in image domains. Cricket shots exhibit strong spatiotemporal locality (bat position relative to body, ball trajectory), making convolutional architectures naturally suited for this task. Additionally, our explicit optical flow integration provides motion-specific inductive bias that attention mechanisms must learn implicitly from data. Synergistic multimodal architecture design The architectural integration of three complementary modalities creates a robust and comprehensive shot recognition system: RGB video processing (3D CNN core) The 3D convolutional architecture with filter progression (32→64→128→256) and strategic pooling (early layers preserve temporal dimension via 1 × 2 × 2 pooling, deeper layers apply full 2 × 2 × 2 reduction) enables hierarchical spatiotemporal feature learning. Global average pooling compresses learned features into compact 256-dimensional representations, preventing overfitting while maintaining discriminative power. The fully connected layers (512→256→7) with dropout regularization (0.4) provide robust classification with minimal parameter overhead. Motion feature integration (optical flow) Farneback optical flow computation provides dense motion field estimation between consecutive frames, capturing bat swing velocity, trajectory curvature, and body movement dynamics. The motion intensity features (19-dimensional vectors representing temporal motion profiles) enable two critical functionalities: (1) implicit augmentation of the 3D CNN’s temporal modeling through motion-aware preprocessing, and (2) explicit shot timing comparison via cosine similarity against professional reference shots, yielding quantitative performance metrics (e.g., 86.79% similarity indicating “Excellent shot timing”). Biomechanical feature extraction (YOLOv8-Pose) Real-time pose estimation (90.3ms inference) provides 17-keypoint skeletal trajectories that encode biomechanical patterns crucial for technique-dependent shots. The lightweight YOLOv8n-pose variant balances accuracy and efficiency, achieving confidence scores of 0.57–0.86 on broadcast footage with variable player scales and partial occlusions. Pose features add interpretability, a critical requirement for coaching applications where understanding why a shot was classified a certain way enhances training effectiveness. Training optimization and generalization The training strategy employed ensures robust convergence and strong generalization: AdamW optimizer with weight decay (0.0001) provides L2 regularization that prevents overfitting during the 50-epoch training process. The adaptive learning rate schedule (initial 0.001, reduced by 0.5 at epochs 6, 20, 33 when validation loss plateaus) enables aggressive early learning followed by fine-grained optimization, achieving validation accuracy of 95.61% that closely matches training accuracy (95.49%), indicating minimal overfitting. The 80 − 20 train-test split (1,021 training, 752 test samples) across a class-balanced dataset (approximately 150 videos per shot category) ensures representative evaluation. Data augmentation (horizontal flipping, brightness variation, spatial cropping) enhances model robustness to camera angle variations and lighting conditions typical in broadcast footage. Batch normalization after each convolutional layer stabilizes training by normalizing layer inputs, while dropout (0.4) in fully connected layers prevents co-adaptation of neurons. This combination enables training on a moderately-sized dataset without requiring massive data volumes typical of deep learning approaches. Practical deployment considerations The system architecture supports immediate deployment in real-world cricket analytics scenarios: Real-time capability YOLOv8-Pose inference (90.3ms) combined with efficient 3D CNN processing enables near-real-time analysis suitable for live match scenarios. The frame extraction strategy (20 uniformly sampled frames per shot) balances temporal coverage with computational efficiency, allowing analysis of 3–5 s shot clips within practical latency constraints for coaching feedback. Quantitative performance metrics The cosine similarity-based shot timing comparison framework provides interpretable performance scores that translate directly into coaching feedback: scores > 0.85 indicate “Excellent shot timing”, 0.75–0.85 suggest “Decent timing, improve consistency”, and < 0.75 highlight specific technical deficiencies. This quantitative assessment democratizes access to professional-grade performance analytics previously limited to elite training facilities with expensive motion capture systems. Scalability and accessibility The system operates on standard broadcast footage without requiring specialized hardware (Vicon systems, IMU sensors, controlled lighting), making it accessible to amateur players and grassroots cricket programs. The reference shot database (7 professional player videos across shot categories) can be expanded to include skill-level-appropriate comparisons, enabling adaptive coaching tailored to player progression. Methodological contributions and scientific rigor Beyond performance metrics, this work advances cricket analytics methodology: Multimodal feature fusion framework The integration architecture demonstrates how complementary modalities (appearance, motion, biomechanics) can be combined effectively for sports action recognition. While some approaches use early fusion (concatenating features before classification), our design employs implicit fusion where modalities serve distinct roles: 3D CNN for primary classification, optical flow for motion comparison, pose for biomechanical interpretability, avoiding dimensionality explosion while maintaining each modality’s specialized contribution. Cricket-specific dataset and benchmark Utilizing the Cricket Shots-IPL 2023 dataset (1,773 videos, 7 categories, class-balanced)21 provides a standardized benchmark for future research. The dataset’s focus on professional IPL footage ensures realistic complexity (variable camera angles, crowd backgrounds, diverse player styles) that tests model robustness beyond controlled laboratory conditions. Reproducible training protocol Complete specification of hyperparameters (learning rate schedule, batch size 16, AdamW with weight decay 0.0001, dropout 0.4, 50 epochs with early stopping), architecture details (filter sizes, pooling strategies, activation functions), and preprocessing steps (75 × 75 frame resolution, 20-frame sampling) enables reproducibility, a critical requirement for advancing scientific understanding in sports analytics. Broader implications and extensions The demonstrated approach has implications extending beyond cricket shot classification: Transfer to related domains The spatiotemporal learning framework is applicable to other bat-and-ball sports (baseball batting, table tennis strokes, tennis shots) where action recognition requires capturing both body positioning and implement (bat/racket) motion dynamics. The multimodal architecture provides a generalizable template for sports action recognition tasks. Enhanced coaching ecosystems Integration with video analysis platforms enables automated highlight generation, technique trend tracking across practice sessions, and data-driven training plan optimization. The quantitative similarity scores facilitate objective progress monitoring, replacing subjective coach assessments with reproducible metrics. Broadcasting and fan engagement Real-time shot classification can augment live broadcasts with automatic annotations, enriching statistical commentary and enabling advanced analytics overlays. The system supports generation of shot heatmaps, player-specific shot distribution analyses, and match situation-specific shot effectiveness metrics that enhance viewer understanding. Talent identification Automated analysis of amateur match footage enables scouts to identify promising players based on shot repertoire diversity, technique quality scores, and consistency metrics, reducing geographical and socioeconomic barriers in talent discovery pipelines. Synthesis and forward outlook The multimodal 3D CNN architecture presented in this work establishes a new state-of-the-art in cricket shot classification through the synergistic integration of spatiotemporal learning, motion-specific features, and biomechanical representations. The 95.61% accuracy, consistent performance across shot categories (F1: 0.93–0.98), and practical deployment characteristics demonstrate both scientific advancement and real-world applicability. The architectural choices, volumetric 3D convolutions over sequential processing, explicit optical flow integration for motion modeling, real-time pose estimation for biomechanical context, represent principled design decisions validated through superior performance compared to established approaches. The training methodology ensures robust generalization, while the quantitative performance comparison framework provides actionable coaching insights. This work demonstrates that intelligent sports analytics systems can achieve professional-grade performance using commodity hardware and broadcast footage, democratizing access to advanced training technologies. The multimodal integration framework, comprehensive experimental validation, and deployment-ready architecture collectively advance the field of cricket analytics while providing a foundation for broader applications in sports action recognition and biomechanical assessment. Conclusion In summary, the cricket shot classification model introduced through this research has proven not only to be very accurate but also to be reliable in the identification and classification of different batting shots. Being that it has the overall accuracy of 95% and strong performance metrics across different shot types, the model is very appropriate for implementation in sports analytics applications such as analyzing batting shot selection and comparing players based on their performance. While there are some misclassifications still, especially for “cut” and “pull” shots, the performance of the model is good enough to extract valuable insights from the data of the players. The future works of this model can include improving the model such that it can overcome these misclassifications too. New features and model architectures can also be used to improve the model performance. The results show that the current model is well trained and generalised so that it can provide proper insights into improving the player’s performance and also to classify the batting shots properly.