A keypoint detection-based approach for real-time aggressive behavior identification in goats

Abstract

This study addresses the existing challenges in animal aggression behavior detection by proposing a real-time recognition method based on keypoint detection. Traditional detection methods based on static whole-frame images are limited in their ability to analyze dynamic processes, often leading to false positives, where non-aggressive behaviors are misclassified as aggression. While algorithms incorporating Long Short-Term Memory (LSTM) networks improve accuracy, they suffer from poor real-time performance due to the need for comprehensive analysis of sequential video data, resulting in a high computational burden and an inability to pinpoint specific aggression events. Furthermore, many existing approaches require artificial interventions during the data collection phase, such as increasing animal density to provoke aggression, which can induce stress responses in the animals and generate data that deviate from natural behaviors, thus failing to accurately reflect authentic social interactions and conflict mechanisms. To overcome these challenges, this study introduces a novel method that detects keypoints on the head and body of goats, analyzing the movement relationships between these keypoints in real-time to identify aggressive behavior without the need for artificial intervention. Experimental results demonstrate that this approach outperforms existing methods in terms of accuracy, recall rate, and false positive rate, while offering superior real-time performance and robustness. Moreover, the proposed method is highly extensible, not only applicable to aggression detection but also capable of analyzing other daily behaviors of goats through keypoint data. To contribute to the advancement of the field of animal behavior studies, we have made publicly available a dataset containing keypoint information of goats, providing valuable resources for future research.