Group-housed pig detection in video surveillance of overhead views using multi-feature template matching

Abstract

In this paper, a novel pig detection approach is proposed for video surveillance based on overhead views in complex group-housed environments. In contrast to traditional approaches that use basic image processing techniques, the proposed algorithm focuses on particular appearance features of pigs and employs the template matching framework to facilitate pig detection. First, to capture the key appearance characteristics of pigs in templates, we represent images as the combination of two complementary descriptors, i.e., dominant orientation templates (DOTs) and brightness ratio templates (BRTs). A DOT establishes relationships with edge features, while a BRT encodes intensity and texture features. Both descriptors are given in binary form to increase the time efficiency of the algorithm. Then, representative pig templates can be obtained from training images through an automatic selection process based on template clustering. Finally, the input image is scanned to compute matching responses for all templates. Thus, objects having similar visual properties as the given templates are considered as detected pigs. Experiments are conducted on the test data extracted from a surveillance video covering 5 days. The experimental results show that our algorithm achieves an average detection rate of 86.8% with a low false alarm rate, which outperforms methods based on threshold segmentation. As a result, the proposed method can effectively and reliably extract pigs in complex scenes including multiple sources of disturbances such as uneven illumination, foreground objects of varying colour, pigs moving slowly, etc.