Early detection of dead broilers in commercial farms using temporal persistence of stationary behavior Authors: Eunji Park, Juwhan Song Published: 2026 ABSTRACT AbstractEarly detection of dead broilers is essential for maintaining productivity and biosecurity in commercial farms; however, dead broilers are difficult to distinguish from live broilers in deep resting states using single-frame images. This study aimed to develop and evaluate an early warning pipeline that detects dead broiler candidates by quantifying the temporal persistence of stationary behavior in broilers aged 1 to 10 days. Top-view CCTV videos were collected from a commercial farm housing approximately 30,000 broilers, and frames were sampled at a rate of 1 frame per minute. Broilers were detected using a YOLO-based object detector, and bounding boxes from the first frame were fixed to extract broiler-level image sequences. A ResNet-BiLSTM classifier categorized each 8-frame sequence as ‘stationary’ or ‘moving’, and consecutive predictions were accumulated to compute stationary duration for each broiler. To set an operational threshold, stationary duration distributions from 93,085 normal observations in the June dataset were analyzed to derive three criteria: the top 1%, top 0.1%, and day-of-age-specific maximum values, which were then evaluated on an independent July dataset. All criteria achieved a recall of 0.95, but precision differed markedly. The maximum stationary duration observed among normal broilers was 91 frames, approximately 91 min. Based on this empirical upper bound, a conservative fixed threshold of 92 frames yielded a precision of 0.7917 and an F1 score of 0.8636. A more conservative threshold of 113 frames increased precision to 0.9048 and the F1-score to 0.9268. When the time interval between exceeding the stationary duration threshold and manual removal was defined as the early detection interval, mean intervals were 479.8 min, approximately 8.0 h, at 92 frames and 458.8 min, approximately 7.6 h, at 113 frames. These results suggest that dead broiler detection can be approached not as a single-frame morphological discrimination problem, but as a temporal persistence problem that evaluates how long a stationary state is maintained. Furthermore, thresholds derived from the behavioral distribution of normal broilers may enable a practical early warning system for detecting dead broilers in commercial broiler farms.Keywords: Broiler farming, Dead broiler detection, Temporal persistence of stationary behavior, Behavior classification, Threshold selection Introduction Rapid and accurate detection of dead broilers is a critical component of maintaining productivity and biosecurity in broiler production systems. Delayed removal of dead broilers can increase secondary risks, including disease transmission and odor generation, thereby compromising farm productivity and biosecurity (Gržinić et al., 2023). In commercial broiler production, first-week mortality (FWM) is widely used as an independent management indicator and accounts for a substantial proportion of total flock mortality. An average FWM of 1.40% and mortality after the first week (MAFW) of 3.05% in Ross 308 broilers were reported, indicating that approximately 31% of total mortality occurred within the first seven days (Forseth et al., 2024). Similarly, analysis of 1,142 production cycles reported a median FWM of 0.95% and total mortality of 3.42%, suggesting that early mortality accounts for approximately 28% of overall losses (Slegers et al., 2025). These findings indicate that mortality during the first week typically represents approximately one-quarter to one-third of total flock mortality, emphasizing the importance of early detection during the initial growth phase. Immediately after placement, chicks undergo physiological adaptation to a new environment. Immature thermoregulation, limited immune competence, and environmental stress increase the risk of early mortality. In practice, the standard method involves farm workers patrolling the poultry house to visually identify and remove dead broilers (Olejnik et al., 2022). However, in high-density rearing environments housing thousands to tens of thousands of broilers, the labor burden is significant, and there is a high probability that dead broilers may be missed or left unattended for extended periods. As time passes, unattended dead broilers may be deformed by the movement and contact of surrounding birds or obscured by litter, making visual identification increasingly difficult. Specifically, once a dead broiler is buried in the litter, farm workers are substantially more likely to miss it, which may compromise farm biosecurity by increasing the risk of disease transmission. Therefore, early detection within the critical window before burial is essential to facilitate worker operations. Early studies on automated broiler monitoring established foundational approaches for detecting inactivity and mortality using invasive sensor-based methods and machine vision techniques (Okinda et al., 2019). Among these approaches, one representative method directly measured individual movement using accelerometers and detected anomalies by exploiting the characteristic flat signal patterns that appear when a bird ceases to move (Bao et al., 2021). While these studies demonstrated that mortality could be determined by quantifying the difference between activity and inactivity, attaching sensors to individual birds presents significant limitations for large-scale application in commercial farms. Non-invasive image-based approaches have also been actively researched. Methods combining thermal and visible light cameras were proposed to detect dead broilers by extracting high temperature living birds from thermal images and removing them from visible light images to isolate dead broilers (Muvva et al., 2018). Subsequently, numerous studies have employed You Only Look Once (YOLO)-based object detection models for dead-broiler detection, including YOLOv3 (Hao et al., 2022), YOLOv5 (Bist et al., 2023), and YOLOv8, YOLOv9, YOLOv10, and YOLOv11 (Bumbálek et al., 2025). Although high performance with a mAP of over 98% has been reported, the challenge lies in the fact that broilers in a deep resting state and actual dead broilers are structurally similar in images (Liu et al., 2021). For instance, transformer-based dual-stream networks have been proposed to integrate detection and segmentation to improve dead-bird identification in farm environments, achieving a mean intersection over union (IoU) of 88% on test data (Ham et al., 2024). However, they remain primarily reliant on single-frame visual information and do not explicitly address the temporal persistence of immobility. Accordingly, recent research has expanded to incorporate temporal information. A previous study proposed a method for screening dead-broiler candidates by calculating the Average Optical Flow Intensity (AOFI), followed by confirming mortality based on how frequently each individual was classified as candidate death across the video (Hao et al., 2025). Another study reported a two-stage approach that first detects stationary regions in CCTV footage and then utilizes a DeiT-Tiny-based classifier to determine if the region is an actual broiler. This study also used prolonged stationary state as a key indicator for mortality determination (Khanal et al., 2025). These studies are significant in that they move beyond simple object detection to utilize the persistence of inactivity along the time axis as a core metric. However, most existing studies have relied on arbitrarily defined thresholds for long-term inactivity, without quantitatively establishing the temporal boundary between stationary behavior and mortality. In one study, chickens were experimentally inoculated with Salmonella and continuously monitored for two weeks using cameras; individuals showing no movement for 30 min or more were flagged as potential infection cases (Kim et al., 2023). While this research demonstrated that prolonged inactivity could serve as a practical early warning signal, the inactivity threshold was pre-defined rather than derived from the actual distribution of stationary behavior in healthy chickens. There is still a lack of research that derives thresholds from the distribution of stationary durations in healthy broilers and validates these thresholds within actual commercial farm environments. Therefore, detecting dead broilers requires an approach that goes beyond single-time-point classification and incorporates temporal persistence. To address this need, we propose an integrated analysis pipeline that combines object detection with time-series behavior classification. First, broilers are detected using YOLO-based object detection, and bounding boxes obtained from the initial frame are fixed to subsequent frames to construct bird-level sequences. Subsequently, a ResNet-BiLSTM-based behavior classification model categorizes each sequence as 'stationary' or 'moving' and consecutive stationary predictions are accumulated to quantify the stationary duration for each individual broiler. Furthermore, we analyze the stationary duration distribution of normal broilers to derive a data-driven threshold for stationary duration, which is then used as a criterion for detecting dead broilers. The main contributions of this study are as follows: 1. We propose a time-series-based dead broiler detection pipeline focusing on the early age period (days 1-10), where mortality rates are high. 1. We propose a time-series-based dead broiler detection pipeline focusing on the early age period (days 1-10), where mortality rates are high. 2. We derive a data-driven stationary duration threshold from the distribution of normal broilers and empirically verify that this threshold achieves a recall of 0.95, precision of 0.7917, and F1-score of 0.8636 in a commercial broiler farm environment. 2. We derive a data-driven stationary duration threshold from the distribution of normal broilers and empirically verify that this threshold achieves a recall of 0.95, precision of 0.7917, and F1-score of 0.8636 in a commercial broiler farm environment. 3. Furthermore, we confirm that applying the established threshold allows the system to detect dead broiler candidates an average of 7-8 hours prior to actual manual removal, providing sufficient lead time for farm workers to intervene. 3. Furthermore, we confirm that applying the established threshold allows the system to detect dead broiler candidates an average of 7-8 hours prior to actual manual removal, providing sufficient lead time for farm workers to intervene. These results suggest that temporal persistence analysis based on data-driven thresholds can provide a more objective basis for dead broiler detection than arbitrarily defined inactivity criteria. This approach may also support the development of an early warning system for timely detection and removal in farm operations. Materials and methods Data collection This study was conducted at a commercial broiler farm housing approximately 30,000 broilers. To collect video data, two IP cameras (IPC675LFW-AX4DUPKC-VG, Uniview) were installed at a height of 4.5 m above the floor inside the poultry house. The cameras were positioned to provide a top-down view, looking vertically at the interior of the house. All collected video footage was connected to a Network Video Recorder (NVR; model: UHD-IN504P, UNDPRO) for stable recording and storage. Video data were collected during two production cycles. The first collection period ranged from June 11 to June 20, 2025, and the second from July 27 to August 5, 2025. Both periods captured the early growth phase from day 1 to day 10 of age. A total of four video datasets were obtained by filming two distinct zones within the house. The specific usage of each dataset is outlined in Table 1. The number of broilers detected within the field of view of each camera ranged from a minimum of 100 to a maximum of over 900. The original video resolution was 2,880 × 1,620 pixels. To improve analysis efficiency, the footage was cropped to 2,800 × 1,440 pixels to exclude the timestamp area. Additionally, frames were extracted at a rate of one frame per minute to reduce the data processing load. Only frames between 03:00 and 21:00, during which lighting conditions remained stable, were selected to construct datasets for statistical analysis, model training, and validation. Table 1Summary of collected video datasets and their usage in the study.Table 1 dummy alt textNoaCamerabPeriodcUsage112025.6.11.∼2025.6.20.YOLO model training andSequence-based behavior classification dataset222025.6.11.∼2025.6.20.Threshold selection set312025.07.27.∼2025.08.05.Threshold-baseddead broiler detection validation set422025.07.27.∼2025.08.05.Threshold-baseddead broiler detection validation setaThe dataset identifier number distinguishing separate data collection sessionsbThe identification number of the specific IP camera unit used for video recordingcThe chronological duration of data acquisition Summary of collected video datasets and their usage in the study. aThe dataset identifier number distinguishing separate data collection sessions bThe identification number of the specific IP camera unit used for video recording cThe chronological duration of data acquisition All AI model training and inference were performed in a hardware environment equipped with an Intel Xeon Gold 5218R CPU @ 2.10 GHz and an NVIDIA A100-PCIE-80GB GPU. YOLO-based detection of broiler individuals To estimate stationary persistence at the individual level, a YOLO-based object detection model was employed to detect broiler individuals and their locations. To construct the training dataset, broiler locations were manually labeled with bounding boxes in a total of 30 original images from the No. 1 dataset, selecting three images for each day of age. Given that a single image contains hundreds of individuals, which can degrade detection accuracy for small objects, and to compensate for the limited quantity of training data, a sliding window technique was applied. The original images were divided into small patches of 320×320 pixels, and a 20% overlap was applied between adjacent windows to prevent the loss of objects located at boundary regions. This approach enhanced detection performance for small objects while effectively expanding the scale of the dataset. The final constructed dataset consisted of 2,274 images, which were partitioned into 1,592 images for training, 342 for validation, and 340 for testing. Upon comparing the YOLOv8 (Varghese and Sambath, 2024) and YOLOv11 models, YOLOv8l demonstrated the most stable performance, achieving a precision of 0.963, a recall of 0.94, and an mAP50-95 of 0.809, across different model sizes, both YOLOv8l and YOLOv11l achieved the highest precision of 0.963, as shown in Table 2. Minimizing false positive detections was crucial since the detected broiler regions were used as inputs for the subsequent stationary behavior classification stage. Because YOLOv8l showed a slightly higher recall than YOLOv11l, it was selected for the subsequent experiments. Table 2Performance comparison of YOLOv8 and YOLOv11 variants on the broiler detection test dataset.Table 2 dummy alt textModelPrecisionaRecallbmAP50-95cYOLOv8n0.9550.9350.797YOLOv8s0.9580.9420.808YOLOv8m0.9610.9410.81YOLOv8l0.9630.940.809YOLOv11n0.950.9360.792YOLOv11s0.9590.9380.804YOLOv11m0.9560.9390.806YOLOv11l0.9630.9370.808aPrecision = The proportion of true positive detections among all positive predictions, calculated as TP / (TP + FP)bRecall = The proportion of true positive detections among all actual ground truth objects, calculated as TP / (TP + FN)cmAP50-95 = Mean Average Precision averaged over Intersection over Union (IoU) thresholds from 0.50 to 0.95 with a step size of 0.05 Performance comparison of YOLOv8 and YOLOv11 variants on the broiler detection test dataset. aPrecision = The proportion of true positive detections among all positive predictions, calculated as TP / (TP + FP) bRecall = The proportion of true positive detections among all actual ground truth objects, calculated as TP / (TP + FN) cmAP50-95 = Mean Average Precision averaged over Intersection over Union (IoU) thresholds from 0.50 to 0.95 with a step size of 0.05 Sequence-based behavior classification model In this study, image sequence datasets were constructed to classify broiler behavior as either stationary or moving based on temporal continuity rather than single-frame information, thereby enabling subsequent estimation of stationary duration distributions. Based on the June No. 1 training data, individual sequences were extracted to construct datasets for the “stationary” and “moving” classes. Behavioral labels were defined as follows: •Stationary: As shown in Fig. 1, this label denotes a state in which the individual remains within the detected area or its original positional range throughout the sequence, without substantial displacement. In high-density rearing environments, this category may include minor tremors or posture changes caused by external factors, such as contact with surrounding individuals or trampling, provided that no substantial positional movement occurs. Stationary: As shown in Fig. 1, this label denotes a state in which the individual remains within the detected area or its original positional range throughout the sequence, without substantial displacement. In high-density rearing environments, this category may include minor tremors or posture changes caused by external factors, such as contact with surrounding individuals or trampling, provided that no substantial positional movement occurs. •Moving: As shown in Fig. 2, this label denotes a state in which the individual clearly moves outside the detected area or shows substantial positional displacement within the sequence. Moving: As shown in Fig. 2, this label denotes a state in which the individual clearly moves outside the detected area or shows substantial positional displacement within the sequence. Examples of image sequences labeled as the ‘Stationary’ class. Examples of image sequences labeled as the ‘Moving’ class. The number of sequences constructed for each class and sequence length is summarized in Table 3. Table 3Composition of the dataset for spatiotemporal behavior classification across varying sequence lengths.Table 3 dummy alt textSequence LengthaMovingbStationarycTraindTesteTotalf31,2001,2001,9204802,40062,0632,0633,3008264,12682,3512,3513,7619414,702aSequence Length = The number of consecutive frames extracted to form a single input sequence for the model (e.g., 3, 6, or 8 frames)bThe number of labeled sequence samples assigned to the ‘Moving’ classcThe number of labeled sequence samples assigned to the ‘Stationary’ classdNumber of data used for trainingeNumber of data used for testingfTotal number of data used for training and testing Composition of the dataset for spatiotemporal behavior classification across varying sequence lengths. aSequence Length = The number of consecutive frames extracted to form a single input sequence for the model (e.g., 3, 6, or 8 frames) bThe number of labeled sequence samples assigned to the ‘Moving’ class cThe number of labeled sequence samples assigned to the ‘Stationary’ class dNumber of data used for training eNumber of data used for testing fTotal number of data used for training and testing A binary classification model was designed to classify individual image sequences into stationary and moving states. Three representative behavior recognition architectures capable of learning spatiotemporal features were employed for training and performance evaluation. Comparative experiments were conducted for a sequence length of T∈{3,6,8} frames using three deep learning model architectures: 3D-CNN (Ji et al., 2013), ResNet-BiLSTM combining a ResNet-18 backbone CNN with BiLSTM (Schuster and Paliwal, 1997), and SlowFast (Feichtenhofer et al., 2019). Each sequence is defined as Eq. (1). where It represents the cropped image of the individual at time point t. All experiments were trained using the AdamW optimizer and CrossEntropyLoss. The dataset was split into an 8:2 ratio for performance evaluation. The experimental results are shown in Table 4. The ResNet-BiLSTM model demonstrated the most stable performance across all sequence lengths, recording a maximum Test Accuracy of 0.9330, particularly at T=8. This indicates that the ResNet-18-based CNN effectively learns frame-level spatial representations, while the Bi-LSTM utilizes bidirectional time-series information to simultaneously reflect past and future contexts, thereby enabling stable modeling of the temporal persistence of stationary behavior in broilers. Table 4Performance comparison of sequence-based behavior classification models across varying sequence lengths.Table 4 dummy alt textModelaSequence LengthbTrain AccuracycTest Accuracyd3DCNN3680.84010.82970.82210.90210.87170.8905ResNet-BiLSTM3680.89320.91300.93030.88960.90920.9330SlowFast3680.73540.71820.71310.85830.80750.7832aThe deep learning architecture employed for binary behavior classificationbThe number of consecutive frames extracted to form a single input sequence for the model (e.g., 3, 6, or 8 frames)cThe accuracy achieved on the training datasetdThe accuracy achieved on the testing dataset Performance comparison of sequence-based behavior classification models across varying sequence lengths. aThe deep learning architecture employed for binary behavior classification bThe number of consecutive frames extracted to form a single input sequence for the model (e.g., 3, 6, or 8 frames) cThe accuracy achieved on the training dataset dThe accuracy achieved on the testing dataset Dead broiler detection pipeline based on temporal persistence of stationary behavior The proposed pipeline for individual broiler analysis is presented in Fig. 3. A sliding window-based sequence construction method was employed to generate individual-level time-series sequences, from which stationary duration was quantified. Letting the chronologically ordered frames be t0,t1,⋯,tn, the analysis was performed repeatedly by setting the frame corresponding to the top of each hour (e.g., 03:00, 04:00, ..., 20:00) from 03:00 to 20:00 as a new starting point. Specifically, the frame at the top of each hour was defined as t0, and stationary duration was accumulated independently using subsequent continuous frames. At each start time t0, YOLO-based object detection was performed to detect broiler bounding boxes. The bounding box coordinates of each detected individual were fixed and applied identically to subsequent frames. Using the initial position as a reference, the same coordinate area was cropped from continuous frames to construct individual sequences. For each individual i, an 8-frame sequence is generated as defined in Eq. (2) where Ii(tk) represents the cropped image of individual i at time point tk, and Si(k) denotes the k-th sliding window sequence. Overview of the proposed pipeline for dead broiler detection based on the temporal persistence of stationary behavior. YOLOv8l is used to detect individual broilers in the initial frame, and the detected bounding boxes are fixed and applied to subsequent frames to construct individual-level 8-frame sliding-window sequences. Each sequence is classified as stationary or moving using the ResNet-BiLSTM model. If the sequence is classified as stationary, the window is shifted by one frame and stationary duration accumulation continues. If the sequence is classified as moving, the accumulation is terminated. The sliding-window process is illustrated for Object 1, and the same procedure is applied independently to all detected individuals. Each sequence was input into a binary behavior classification model that extracts spatial features via a ResNet-based CNN and learns temporal dependencies via Bi-LSTM. The model classified each sequence as either ‘Stationary’ or ‘Moving’. If the first 8-frame sequence for a specific individual was classified as moving, stationary duration accumulation was immediately terminated, and Li was set to 0. If the sequence was classified as stationary, the window was shifted by one frame to construct the next 8-frame sequence, and classification was performed repeatedly. This process continued only while consecutive sequences were classified as stationary. The stationary state was considered to have ended when a sequence was classified as moving. Through this process, the number of consecutive stationary windows from the initial sequence, Li was defined for each individual i, as shown in Eq. (3). where yj(i) denotes the behavior classification result of individual i for the j-th 8-frame sliding-window sequence Si(j), and Li represents the number of consecutive windows classified as stationary from the initial sequence. Rather than using the consecutive stationary window count directly, the stationary duration was defined as the corresponding number of image frames, as shown in Eq. (4). This formulation reflects the sliding-window structure in which the first stationary window represents a minimum stationary duration of 8 frames. Each additional overlapping stationary window, shifted by one frame, extends the estimated stationary duration by one frame. In other words, if Li≥1 consecutive stationary windows are detected, the corresponding stationary duration is (Li+7) frames. If no stationary window is detected, the stationary duration is set to 0. Because the frames were sampled at 1-min intervals, this value also corresponds to approximately (Li+7) minutes. Dead broilers typically exhibit the characteristic of maintaining a stationary state in a similar posture at the same location for an extended period until removal. Accordingly, an individual was classified as a dead broiler candidate when its stationary duration exceeded the predefined threshold, as shown in Eq. (5). where θ denotes the stationary-duration threshold, which was set to the maximum stationary duration observed in normal broilers. An example of the proposed pipeline output is shown in Fig. 4 The individual detected at the initial time point remained in the same location with a similar posture for 227 consecutive frames, and its stationary duration was accumulated accordingly. This example illustrates how the proposed pipeline quantifies the temporal persistence of stationary behavior and stops accumulation when movement or removal is detected. Fig. 5 further illustrates this termination process at the sequence level. The 8-frame sequence at T=227 was classified as stationary, whereas the subsequent 8-frame sequence at T=228 was classified as moving; therefore, stationary duration accumulation was terminated at this transition point. Representative example of stationary duration accumulation and termination in the proposed pipeline. (a) The target individual remained at the same location with a similar posture, resulting in continuous accumulation of stationary duration up to T = 227. (b) After removal or displacement of the individual, the corresponding image sequence was classified as 'Moving', and the stationary state was terminated. Example of the transition from stationary to moving classification in the proposed sliding-window pipeline. The upper row shows the 8-frame sequence at T = 227, which was classified as 'Stationary'. The lower row shows the 8-frame sequence at T = 228, which was classified as 'Moving'. This transition indicates the termination of stationary duration accumulation. Age-specific threshold derivation using normal broiler stationary duration distributions To establish the threshold, Dataset No. 2 (refer to Table 1) was used to analyze the distribution of Tstationary for normal broilers. Individuals confirmed as dead broilers were excluded from the statistical analysis, and the distribution was estimated based on 93,085 observations of normal broilers detected in the footage. This exclusion was necessary because including dead broilers would overestimate the duration of prolonged stationary states, potentially distorting the estimation of the upper bound of the stationary duration distribution for normal behavior. If a Tstationary value was calculated for the same individual across multiple time slots, only the Tstationary value from the initial detection was used as the representative value for the analysis. Therefore, by estimating the distribution of stationary duration solely for normal broilers, we aimed to define the range of physiologically possible resting behaviors. Among all normal observations, 6,721 (approximately 7.2%) exhibited a continuous stationary state lasting at least 8 frames. This indicates that stationary behavior lasting at least 8 frames, approximately 8 minutes, was observed only in a limited subset of normal observations. Because observations without any stationary sequence have no measurable stationary duration beyond the initial classification, threshold derivation for dead broiler candidate detection was performed using only the subset of normal observations in which at least one stationary sequence was detected. The distribution of stationary duration in normal broilers allows for the estimation of the upper bound of physiologically normal resting behavior; stationary states that exceed this upper bound can be interpreted as anomalous behaviors outside the general physiological resting range. Since dead broilers typically maintain a stationary state in a similar posture at the same location for an extended period until removal by the farmer, continuous stationary states exceeding the upper bound of normal stationary duration provide a rational basis for classification as dead broiler candidates. As shown in Fig. 6, the distribution of initial 8-frame sequence classifications showed that most normal observations were classified as moving, whereas only a limited proportion exhibited an initial stationary sequence. The proportion of observations with an initial stationary sequence was relatively high during the early rearing period and gradually decreased with age. This trend may reflect the frequent occurrence of prolonged resting behavior during the early adaptation phase, followed by increased activity, feeding, and exploratory behavior as the broilers grew. As shown in Fig. 7, Fig. 8, three statistical criteria were derived for each day of age: the top 1%, the top 0.1%, and the maximum value. The analysis results indicated that the maximum number of stationary frames was relatively high during the early ages, particularly from days 1 to 5, with the highest value of 91 frames recorded on day 3. This can be attributed to the greater tendency of young chicks to remain inactive for extended periods during the early adaptation stage. The top 0.1% threshold during this period was also relatively high, ranging from 49.8 to 78.9 frames. Conversely, the frequency of prolonged stationary states decreased sharply as the broilers reached later days of age. The decrease in prolonged stationary states can be attributed to increased activity accompanying growth, a higher frequency of interactions among individuals, and the activation of feeding and movement behaviors. Daily distribution of initial sequence classifications in normal broilers from day 1 to day 10. The stacked bar chart shows the proportion of normal broiler observations in which the first 8-frame sequence was classified as moving, indicating no initial stationary sequence (blue), and those in which the first 8-frame sequence was classified as stationary, indicating the presence of an initial stationary sequence (orange). The numbers above the bars indicate the total number of normal broiler observations for each day of age. Age-specific distributions of stationary duration in normal broilers. Each histogram shows the frequency of stationary duration values for each day of age on a logarithmic scale. Vertical lines indicate the day-specific statistical thresholds derived from the corresponding distribution: the top 1% value, the top 0.1% value, and the daily maximum value. Statistical stationary duration thresholds (Tstationary) across broiler days of age. Evaluation metrics for dead broiler detection The ground truth (GT) for dead broilers was established by the researchers through manual review of the video data based on the following two criteria: 1. Cases in which the farm worker was clearly observed removing the dead broiler in the recorded footage. 1. Cases in which the farm worker was clearly observed removing the dead broiler in the recorded footage. 2. Cases in which an individual was obscured by litter and was not immediately removed by the farm worker, but was determined to be a dead broiler after continuous visual monitoring of the footage for more than 24 hours. Fig. 9 illustrates a chronological sequence of scenes in which the same individual remains stationary at the same location for over 24 hours. Scenes in which surrounding individuals repeatedly passed over or stepped on the subject were observed, showing an instance in which the dead broiler was gradually covered by litter and was not immediately collected by the farm worker. 2. Cases in which an individual was obscured by litter and was not immediately removed by the farm worker, but was determined to be a dead broiler after continuous visual monitoring of the footage for more than 24 hours. Fig. 9 illustrates a chronological sequence of scenes in which the same individual remains stationary at the same location for over 24 hours. Scenes in which surrounding individuals repeatedly passed over or stepped on the subject were observed, showing an instance in which the dead broiler was gradually covered by litter and was not immediately collected by the farm worker. Chronological sequence of a dead broiler maintaining a stationary state at a fixed location for more than 24 hours. The sequence captures repeated instances of surrounding broilers trampling or passing over the subject. Consequently, the dead broiler becomes progressively obscured by litter, illustrating a scenario in which immediate visual detection and removal by farm workers are hindered due to occlusion. To evaluate the dead broiler detection performance of the proposed system and the threshold-setting strategy, precision, recall, and F1-score were used as evaluation metrics. Individuals with a stationary duration Tstationary(i) greater than or equal to the threshold θ were predicted as dead broiler candidates, and a case in which the intersection over union (IoU) between the predicted bounding box and the ground truth (GT) bounding box was 0.3 or higher were defined as a true positive (TP). Cases in which the IoU was less than 0.3 or no corresponding GT existed were classified as false positives (FP). Additionally, GT dead broilers that were not matched with any predicted bounding box with an IoU of 0.3 or higher were defined as false negatives (FN). IoU was defined as the ratio of the overlapping area between the predicted bounding box and the GT bounding box and was calculated as shown in Eq. (6). Precision, recall, and F1-score were calculated as Eq. (7), Eq. (8), and Eq. (9), respectively. Results and discussion Dead broiler detection performance based on age-specific statistical thresholds The thresholds were evaluated by applying the three statistical criteria derived from the June data, namely the age-specific top 1% value, top 0.1% value, and maximum value, to Datasets 3 and 4 (refer to Table 1), which were not used for training. The evaluation results are shown in Table 5. All three criteria achieved a recall of 0.95. In terms of precision, however, substantial differences were observed depending on the threshold-setting method. The precision based on the top 1% criterion was very low at 0.1044, indicating that the top 1% threshold failed to sufficiently exclude the general resting behaviors of normal broilers. When the age-specific top 0.1% threshold was applied, precision improved to 0.3455, but cases of prolonged resting behavior in normal broilers were still included. Setting the threshold based on the age-specific maximum value resulted in a further improvement in precision to 0.4043; however, limitations still remained in completely distinguishing prolonged normal resting behavior from dead states. Because stationary duration can be influenced not only by age but also by farm-specific conditions, including the rearing environment and chick health and adaptation status, thresholds determined separately for each day of age may be biased when derived from a single rearing cycle. Therefore, rather than determining separate age-specific thresholds, we further evaluated dead broiler detection performance using population-level thresholds derived from the entire set of normal broilerobservations. Table 5Dead broiler detection performance based on age-specific statistical thresholds.Table 5 dummy alt textThresholdRecallPrecisionF1-scoreTop 1%0.95000.10440.1881Top 0.1%0.95000.34550.5067Max0.95000.40430.5672 Dead broiler detection performance based on age-specific statistical thresholds. Dead broiler detection performance based on population-level statistical thresholds The results of analyzing the stationary duration distribution for detection cases classified as normal broiler observations in the June data are presented in Fig. 10. Based on the stationary duration distribution of the entire normal population, the top 1% value, top 0.1% value, and observed maximum value were derived as 39.00, 64.28, and 91.00 frames, respectively. Among these values, 91 frames represents the empirical maximum stationary duration observed in the large-scale normal population and can be interpreted as the empirical upper bound of long-term resting behavior that can occur in normal broilers. Accordingly, dead broiler detection performance was analyzed by incrementally increasing fixed thresholds within the range of 40-120 frames, including 40 frames, 65 frames, and 92 frames, which exceed the top 1%, top 0.1%, and maximum-value criteria, respectively. The results are shown in Fig. 11. As shown in Table 6, recall remained constant at 0.95 across all thresholds, whereas precision improved monotonically as the threshold increased. Specifically, precision increased from 0.1863 at 40 frames to 0.6786 at 65 frames and further to 0.7917 at 92 frames, with corresponding F1-scores of 0.3115, 0.7917, and 0.8636, respectively. When the threshold was increased to 113 frames, precision rose to 0.9048, and the F1-score reached its highest value at 0.9268. This indicates that false positives can be further reduced while maintaining a recall of 0.95. Population-level distribution of stationary duration in normal broiler observations. The histogram shows the frequency of stationary duration values pooled across all normal observation population on a logarithmic scale. Vertical lines indicate the statistical thresholds derived from the entire normal observation population: the top 1% value (39 frames), the top 0.1% value (64 frames), and the empirical maximum value (91 frames). Trends in dead broiler detection performance metrics according to variations in the stationary duration threshold from 40 to 120 frames. The graph illustrates that, while recall remains constant at 0.95, precision and F1-score increase monotonically as the threshold increases, reaching their highest values at the conservative threshold of 113 frames. Table 6Dead broiler detection performance based on population-level statistical thresholds.Table 6 dummy alt textThresholdFramesRecallPrecisionF1-scoreTop 1%400.95000.18630.3115Top 0.1%650.95000.67860.7917Max920.95000.79170.8636Best F1-score1130.95000.90480.9268 Dead broiler detection performance based on population-level statistical thresholds. However, as the threshold increases, alarm generation may be delayed, resulting in a trade-off with the research objective of early detection. Therefore, this study proposes 92 frames, which exceeds the empirical maximum stationary duration observed in normal broilers, as the baseline operational threshold; however, selecting a more conservative threshold, such as 113 frames, may also be feasible depending on operational objectives. Analysis of missed detections All false negative cases, that is, missed detections, occurred because the individuals were not detected during the YOLO-based object detection stage due to occlusion; consequently, adjusting the threshold had no impact on these cases. Examples of such missed detections are presented in Fig. 12. Visualization of dead broiler detection results illustrating a false negative case caused by occlusion. Green bounding boxes represent the ground truth, and red bounding boxes indicate the model predictions. The figure highlights an instance in which a dead broiler was occluded by a feeder line, preventing initial detection by the YOLO model and resulting in a missed detection. Analysis of early detection lead time before manual dead broiler removal To quantify the operational validity of the proposed system, the time difference between the moment when the system identified a stationary state exceeding the threshold and the moment when the worker actually removed the dead broiler was defined as the early detection lead time. As shown in Fig. 13, when the 92-frame threshold was applied, the mean early detection lead time was 479.8 minutes (approximately 8.0 hours), ranging from 87 to 1,890 minutes. When the threshold was increased to 113 frames, the mean early detection lead time decreased slightly to 458.8 minutes (approximately 7.6 hours), ranging from 66 to 1,869 minutes. These results indicate that both thresholds provided an average lead time of more than 7 hours for intervention. Therefore, the threshold can be selected according to operational priorities, considering the trade-off between earlier detection and improved detection accuracy. Early detection lead time distributions for the 92-frame and 113-frame thresholds. The lead time was defined as the interval between system detection and manual removal of the dead broiler. Boxplots show the distribution of lead times for each threshold, with mean values of 479.8 min for θ=92 and 458.8 min for θ=113. Conclusions This study redefined the problem of dead broiler detection in commercial broiler farms from a single-frame-based stationary classification issue to a problem of analyzing temporal persistence along the time axis. In high-density rearing environments, dead broilers and normal broilers in a deep resting state appear structurally similar in single images; thus, it is difficult to accurately distinguish dead broilers from resting normal broilers using only simple object detection-based approaches. Therefore, this study proposed an integrated pipeline that quantifies stationary duration at the individual level by combining YOLO-based object detection with a ResNet-BiLSTM-based time-series behavior classification model. We analyzed the stationary duration distribution of normal individuals to derive an empirical upper bound and presented a threshold-setting strategy that defines continuous stationary states exceeding this bound as dead broiler candidates. The proposed threshold (θ=92) achieved a recall of 0.95, precision of 0.7917, and an F1-score of 0.8636 on an independent dataset. When a more conservative threshold (θ=113) was applied, precision and F1-score improved to 0.9048 and 0.9268, respectively. Furthermore, the proposed system detected dead broiler candidates an average of 7-8 hours before the actual removal time, providing sufficient response time for farm workers to intervene in advance. The results suggest that early detection can be achieved while minimizing false positives by defining dead broiler candidates based on the statistical upper bound of normal behavior. In practical farm management, the proposed pipeline can provide early alerts together with the estimated locations of dead broiler candidates before routine removal by farm workers. Because detected candidates can be represented as bounding-box coordinates within the camera field of view, farm workers can more easily identify locations that require inspection and removal. This may also help workers recover dead broilers that could be missed by visual observation alone due to litter coverage or occlusion. In addition, the system could be extended to record the number and spatial distribution of detected dead broiler candidates over time, thereby supporting mortality monitoring and timely removal in commercial broiler farms. Future research should focus on verifying the generalizability of the thresholds through long-term data accumulation, mitigating occlusion issues through multi-camera or tracking-based approaches, and expanding the system to 24-hour monitoring under nighttime conditions. These advancements are expected to contribute to the establishment of automated health monitoring systems in commercial broiler farms. Ethics approval This study analyzed video data collected in a commercial broiler farm and did not involve any experimental manipulation or invasive procedures. All dead broilers considered in this study resulted solely from natural mortality occurring under standard commercial farming conditions. Funding This research was supported by Korea Institute of Planning and Evaluation for Technology in Food, Agriculture and Forestry (IPET) and Korea Smart Farm R&D Foundation (KosFarm) through Smart Farm Innovation Technology Development Program, funded by Ministry of Agriculture, Food and Rural Affairs (MAFRA) and Ministry of Science and ICT (MSIT), Rural Development Administration (RDA), grant number RS-2025-02216818. Availability of data and materials As the datasets were generated as part of a funded research project, access may be granted by the corresponding author after the project’s completion upon reasonable research request and appropriate justification. Declaration of AI and AI-assisted technologies in the writing process During the preparation of this work, the authors used ChatGPT, an AI-assisted writing tool, to improve clarity and English grammar. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article. CRediT authorship contribution statement Eunji Park: Writing – original draft, Validation, Software, Methodology, Data curation. Juwhan Song: Writing – review & editing, Supervision. Disclosures The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Footnotes Scientific section: Animal Well-Being and Behavior Contributor Information Eunji Park, Email: jijiji1207@jj.ac.kr. Juwhan Song, Email: jwsong@jj.ac.kr. References Associated Data Data Availability Statement As the datasets were generated as part of a funded research project, access may be granted by the corresponding author after the project’s completion upon reasonable research request and appropriate justification.