Automatic recognition of isolated piglet outliers based on multi-object tracking

Early detection and timely intervention of isolated piglet outliers are crucial in promptly identifying potential illnesses and ensuring animal welfare. This study proposes a comprehensive framework, Multi-Object Tracking Isolated Outlier Detection (MOTIOD), to identify persistent isolated piglet outliers in complex pig farm environments. The framework consists of a multi-object tracking (MOT) module and an outlier detection module. The MOT module, employing a transformer-based TransTrack approach, develops dark region-adaptive IoU (Intersection over Union) matching strategy and piglet ID constraint allocation strategy to address common challenges in pig farms, such as variable lighting conditions and occlusions. The outlier detection module presents a novel Spatiotemporal Dynamics-Based Outlier Detection (SD-OD) algorithm, which considers the duration of isolated outlier behaviour and the spatial information of piglet interactions with their sow and littermates. Experimental results validate the MOTIOD framework demonstrates significant performance in low visibility continuous tracking, outperforming state-of-the-art methods such as Yolov5s+DeepSORT and CenterTrack with a MOTA of 98.8 % and an IDF1 of 99.0 %, and effectively identifies isolated outliers with a 92.9 % precision rate in outlier detection. This research introduces a valuable approach for the early identification of potential health risks in piglets, enhancing the proactive management of livestock health, and potentially contributing to the sustainable advancement of precision livestock farming.