Robust Animal Tracking and Stereotypical Behavior Detection Under Real Environment Using Temporal Averaging Background Subtraction

In ethology research, there have been growing interests in using machine learning method to detect animals and analyze their behaviors, especially from video data. However, behavior analysis is still challenging in the outdoor environment because of the dynamic background and sudden illumination changes. Instead of the previous laboratory setting, we aimed to perform animal behavior analysis outdoors. Specifically, our target of detection and behavior analysis is a polar bear captured by a security camera in a zoo. We focus on analyzing stereotypical behavior, which is critical for understanding the psychological stress of animals. For detection and analysis, we proposed a method that includes background extraction, object detection, and repeating pattern detection for stereotypical behavior detection based on the compression ratio of the detected bear’s location sequences under serialization. Our experimental result shows our method could provide accurate detection (98.3%AP50) and behavior recognition (Accuracy 90.6%) while maintaining high robustness to various noises.