Automatically detecting stereotyped behaviors from surveillance videos of zoos and animal reserves for captive animal welfare

Animal welfare plays the importance role in protecting the ecological environment, developing the animal industry, and regulating the human-animal relationship and maintaining the balance of species and ecosystems. Unfortunately, captive animals often suffer from poor welfare due to their long-term confinement in small and narrow environment, resulting in depressive and stereotyped behaviors such as repetitive running or pacing. Early detection and intervention of these behaviors contribute to improve the captive animal welfare. However, the traditional methods on manually observing and analyzing animal behaviors are labor-intensive and require specialized personnel. In recent years, many zoos and animal reserves are equipped with video surveillance systems, which can be conveniently used for animal behaviour analysis, but most research on the animal behaviour is reported for domestic animal. In this paper, we propose the novel method utilizing the videos from surveillance systems to automatically track and recognize animals’ stereotyped behaviors. Our approach employs a novel target tracking method based on Siamese relation network, which can accurately track target animals in complex environments and distinguish different individuals. We then identify the stereotyped behaviors by counting the cycle number that an animal repeats similar motion trajectories, such as repeatedly running around a similar circular trajectory. The experiments are conducted on the video datasets including three kinds of moving trajectories. The experimental results show that can accurately detect the cycle number of the repeated trajectories to recognized wild animal stereotypes.