Early detection of dead broilers in commercial farms using temporal persistence of stationary behavior

Early 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.