Activity Recognition from Smart Insole Sensor Data Using a Circular Dilated CNN Yanhua Zhao Abstract Smart insoles equipped with pressure sensors, accelerometers, and gyroscopes offer a non-intrusive means of monitoring human gait and posture. We present an activity classification system based on a circular dilated convolutional neural network (CDCNN) that processes multi-modal time-series data from such insoles. The model operates on 160-frame windows with 24 channels (18 pressure, 3 accelerometer, 3 gyroscope axes), achieving 86.42% test accuracy in a subject-independent evaluation on a four-class task (Standing, Walking, Sitting, Tandem), compared with 87.83% for an extreme gradient-boosted tree (XGBoost) model trained on flattened data. Permutation feature importance reveals that inertial sensors (accelerometer and gyroscope) contribute substantially to discrimination. The approach is suitable for embedded deployment and real-time inference. 1 Introduction Wearable sensors embedded in footwear enable continuous monitoring of gait, balance, and posture without restricting natural movement [ 1 ] . Applications range from fall-risk assessment in older adults [ 2 ] to sports performance analysis and rehabilitation. Smart insoles typically integrate pressure mats with inertial measurement units (IMUs), yielding high-dimensional time-series that capture both ground-reaction patterns and foot motion. Existing work on smart insole systems has explored hand-crafted features and classical machine learning models for gait event detection and activity recognition, for example using time- and frequency-domain pressure statistics combined with support vector machines or random forests [ 3 , 4 ] . More recently, deep learning has become the dominant paradigm for human activity recognition from wearable sensors, with convolutional and recurrent architectures applied to accelerometer and gyroscope streams collected from smartphones and body-worn devices [ 5 ] . However, many of these approaches focus on inertial data alone, or treat pressure and inertial modalities independently, leaving open the question of how best to fuse all available channels from a smart insole in a unified model. Beyond wearables, radar [ 6 ] and vision-based systems [ 7 ] have also been proposed for activity recognition and gait analysis. Camera setups can capture rich kinematic information but raise privacy concerns, depend on lighting conditions and line-of-sight, and are difficult to deploy outside controlled environments. Radar sensors can operate in low light and offer some through-obstacle capability, yet typically require fixed infrastructure, careful calibration, and sophisticated signal processing, and they still observe subjects from a distance rather than at the point of contact with the ground. In contrast, smart insoles provide subject-specific, viewpoint-invariant, and inherently privacy-preserving measurements that naturally integrate into everyday footwear and can be used continuously in free-living conditions. In this work we focus on end-to-end learning directly from the raw multi-modal insole signals. We propose a circular dilated 1D convolutional neural network (CDCNN) that treats each time window as a sequence over the feature dimension, applying dilated convolutions with circular padding to exploit temporal dependencies without losing resolution. The model is trained on a dataset of 14,748 labeled windows and evaluated with a held-out test set. We further employ permutation feature importance to quantify the contribution of each sensor channel and to provide insight into the relative roles of pressure and inertial measurements. 2 Method 2.1 Dataset and Preprocessing The dataset is compiled from recordings of multiple subjects performing four activities: Standing, Walking, Sitting, and Tandem stance [ 8 ] . Each sample is a fixed-length window of 160 time steps with 24 features per frame: – 18 pressure sensors (pressure_0 to pressure_17) sampled per time step – 3-axis accelerometer (accel_x, accel_y, accel_z) in units of g – 3-axis gyroscope (gyro_x, gyro_y, gyro_z) Input matrices are reshaped to (N,T,F)=(N,160,24) , where N is the batch size, T the number of frames, and F the number of channels. Class distribution is approximately balanced (Standing: 5559; Walking: 5401; Sitting: 5066; Tandem: 5043 samples). 2.2 CDCNN Architecture Compared with standard CNNs or recurrent architectures, the CDCNN offers several advantages for smart insole data. Dilated convolutions expand the effective temporal receptive field exponentially with depth while keeping the number of parameters small, enabling the model to capture both short-term foot contacts and longer-term gait patterns within a 160-frame window. Circular padding avoids boundary artifacts at the start of each segment, which is important when activities do not align perfectly with window borders. In addition, the purely convolutional design is fully parallelizable across time, leading to lower inference latency than RNN-based models and making the approach well suited for real-time deployment on embedded hardware inside the insole. Each input sample is a window of T=160 time steps with F=24 sensor channels (pressure, accelerometer, gyroscope). Features then are sent through a stack of circular dilated convolutional blocks that aggregate information over increasingly longer time scales. A final global average pooling over time produces a compact 64-dimensional representation for each window, which is fed to a linear classifier that outputs logits for the four activity classes. The architecture comprises: Circular dilated convolution blocks. Each block consists of a 1D convolution with dilation d=2^{i} for layer i , circular padding to preserve sequence length, batch normalization, and ReLU. Four such blocks use dilations 1, 2, 4, and 8, increasing receptive field without pooling. Hidden channels are set to 64; kernel size is 3. Dropout (0.2) is applied after each block. Classification head. A global average pooling layer reduces the temporal dimension to 1, followed by a fully connected layer mapping 64 hidden units to the number of classes (4). Cross-entropy loss is minimized with the Adam optimizer (learning rate 0.01). Early stopping with patience 20 on validation accuracy prevents overfitting. Figure 1: Class distribution of the compiled smart insole dataset. 2.3 Permutation Feature Importance To assess which channels contribute most to classification, we apply permutation importance [ 9 ] . Let \mathcal{D}_{\text{test}}=\{(x_{i},y_{i})\}_{i=1}^{N} denote the test set and M the trained CDCNN. The baseline accuracy is A_{\text{base}}=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}\big[M(x_{i})=y_{i}\big]. (1) For a given feature (channel) f , we construct a permuted dataset \tilde{\mathcal{D}}^{(f)} by randomly shuffling the values of feature f across samples while keeping all other features unchanged. The corresponding accuracy is Figure 2: Example tri-axial accelerometer window for a single activity segment. A_{f}^{\text{perm}}=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}\big[M(\tilde{x}_{i}^{(f)})=y_{i}\big], (2) where \tilde{x}_{i}^{(f)} denotes the input with the permuted feature f . The permutation importance of feature f is defined as the accuracy drop: I_{f}=A_{\text{base}}-A_{f}^{\text{perm}}. (3) Figure 3: Confusion matrix for CDCNN model on the held-out test set. Larger values of I_{f} indicate a stronger contribution of feature f to the classifier’s performance. This procedure is applied to every channel over the held-out test set. Table 1: Number of samples per subject and class. Subject ID Sitting Standing Tandem Walking Total \@BTrule []1 0 404 0 426 830 10 0 392 0 412 804 12 0 430 0 358 788 13 0 358 0 360 718 14 396 396 0 147 939 15 566 398 590 406 1960 16 548 368 215 370 1501 17 504 153 564 418 1639 18 342 0 0 0 342 19 592 388 594 240 1814 22 408 390 510 386 1694 23 330 372 546 368 1616 24 230 360 514 366 1470 30 384 382 550 410 1726 31 442 372 420 406 1640 32 324 396 540 328 1588 3 Results 3.1 Dataset Overview To enforce subject independence between training and evaluation, subjects are partitioned into disjoint sets using a fixed random seed. Table 2 lists the subject IDs assigned to each split and the resulting number of windows. Table 2: Subject-wise train/validation/test split and sample counts. Split Subject IDs # samples \@BTrule []Train 1, 10, 12, 14, 15, 18, 19, 23, 30, 31, 32 14,047 Val 17, 22 3,333 Test 13, 16, 24 3,689 The dataset contains recordings of multiple subjects performing the four target activities. Fig. 1 summarizes the class distribution, which is relatively balanced across Standing, Walking, Sitting, and Tandem, reducing the risk of bias toward any single activity. Table 1 details the number of windows per subject and per class, highlighting that some subjects contribute more samples and that the distribution of activities varies slightly across subjects. Fig. 2 illustrates a representative tri-axial accelerometer window, showing distinct patterns for the dynamic phases of the motion. 3.2 Classification Performance Training was run for up to 300 epochs with early stopping on the subject-wise validation set. On the held-out subject-wise test set, the CDCNN achieved a test accuracy of 86.42% . For comparison, an XGBoost model trained on data of the same subject-wise split reached a higher test accuracy of 87.83% , indicating that gradient-boosted trees remain a strong reference on relatively low-dimensional, fixed-length tabular representations. One likely reason for the performance gap is that XGBoost can exploit rich non-linear interactions between all time-step features in the flattened vector, whereas the CDCNN is constrained by its local convolutional filters and may not be fully tuned to the limited dataset size and cross-subject variability. In addition, the CDCNN hyperparameters (depth, hidden width, regularization) were only coarsely tuned, whereas XGBoost is known to be robust with moderate tuning. Despite the slightly lower accuracy, the CDCNN offers several potential advantages. First, it operates directly on the sequence structure, enabling extensions to variable-length windows or online streaming without re-flattening. Second, the temporal feature maps of the CDCNN can be inspected or constrained (e.g., via attention or class activation mapping) to provide time-resolved interpretability that is harder to obtain from ensembles of decision trees. The model converges quickly; validation accuracy exceeds 80% within a few epochs. Figure 4: Permutation feature importance by channel over the test set. Figure 5: Permutation feature importance by channel over the test set for the XGBoost. 3.3 Feature Importance Permutation importance analysis of the CDCNN shows that inertial channels (accelerometer and gyroscope axes) have consistently higher importance than many individual pressure sensors. This suggests that motion dynamics are highly informative for discriminating the four activities. Among pressure sensors, spatial patterns vary; the heel and toe regions often show elevated importance. The combined pressure modality remains essential for characterizing stance and contact patterns. Fig. 4 shows the per-channel permutation importance for the CDCNN, highlighting the relative impact of each sensor stream on overall accuracy. For comparison, we also compute permutation importance for the XGBoost on the same subject-wise test set (Fig. 5 ), observing broadly similar rankings but with slightly higher relative importance assigned to some pressure channels, reflecting the different inductive biases of tree ensembles versus convolutional networks. 4 Conclusion We presented a CDCNN-based activity recognition pipeline for smart insole data that fuses pressure and inertial modalities. On a challenging subject-independent split (no subject overlap between train, validation, and test sets), the model achieves 86.42% test accuracy with a simple, interpretable architecture and performs competitively with an XGBoost model (87.83%) trained on flattened feature data. Permutation feature importance indicates that accelerometer and gyroscope channels contribute substantially, complementing pressure-based cues. The approach is suitable for real-time deployment given its low computational footprint. 5 Outlook Future work will address several directions. First, subject-independent evaluation (leave-one-subject-out) will better assess generalization to new users. Second, extending the model to recognize finer activities (e.g., stair ascent/descent, turning) and continuous gait events (stance, swing phases) would increase clinical utility. Third, lightweight model variants and pruning could enable on-device inference on resource-constrained insoles. Finally, incorporating temporal smoothing or recurrent layers may improve robustness to label noise and transient misclassifications.