Enhanced camera-based individual pig detection and tracking for smart pig farms

Abstract

Negative social interactions are harmful for animal health and welfare. It is increasingly important to employ a continuous and effective monitoring system for detecting and tracking individual animals in large-scale farms. Such a system can provide timely alarms for farmers to intervene when damaging behavior occurs. Deep learning combined with camera-based monitoring is currently arising in agriculture. In this work, deep neural networks are employed to assist individual pig detection and tracking, which enables further analyzing behavior at the individual pig level. First, three state-of-the-art deep learning-based Multi-Object Tracking (MOT) methods are investigated, namely Joint Detection and Embedding (JDE), FairMOT, and YOLOv5s with DeepSORT. All models facilitate automated and continuous individual detection and tracking. Second, weighted-association algorithms are proposed for each MOT method, in order to optimize the object re-identification (re-ID), and improve the individual animal-tracking performance, especially for reducing the number of identity switches. The proposed weighted-association methods are evaluated on a large manually annotated pig dataset, and compared with the state-of-the-art methods. FairMOT with the proposed weighted association achieves the highest IDF1, the least number of identity switches, and the fastest execution rate. YOLOv5s with DeepSORT results in the highest MOTA and MOTP tracking metrics. These methods show high accuracy and robustness for individual pig tracking, and are promising candidates for continuous multi-object tracking for real use in commercial farms.