Occlusion-resistant locomotion analysis of piglets using amodal instance segmentation

Locomotion of piglets is a critical indicator of their growth, health, and welfare status; thus it is of utmost importance to automate the analysis of piglet locomotion, particularly during the early lactation periods. An intersection over unit- (IOU-) and contour-based tracking method is proposed to automate the locomotion analysis for piglets in farrowing pens. In the first step, an anchor-free deep learning network is employed in amodal instance segmentation of individual piglets. Then a novel attention graph convolution-based structure is used to distil element-wise features within the detected piglets. The distilled features are further encoded by a graph convolutional network. In the output features, pixels selected by a selection strategy derive features for a real-value pixel using 4-point nearest neighbor bilinear interpolation. Thus, a higherresolution segmentation is predicted in a coarse-to-fine fashion. In the second step, a regional matching strategy is adopted synchronously for initial tracking. The Hungarian algorithm is then used to optimize initial tracking trajectories determined by contour features and IOUs. In the experiment, our method produced crisp amodal instance segmentation, whilst also achieving favorable tracking performance under occlusion. The tracking results demonstrated an IDF1 score of 95.7% and an MOTA of 97.1% in short video clips. With the high-accuracy occlusion-resistant segmentation and tracking results, our computer vision-based piglet tracking method may aid automated piglet locomotion and behavior analysis.