Abstract: The modern pig farming industry tends to scale up and intensify, and the aggressive behavior among gregarious pigs occurs frequently, which seriously affects their health and production, and economic benefit. An end-to-end anchor-free temporal localization framework is proposed to automatically detect aggressive behavior among gregarious pigs, its occurrence and their time periods by means of surveillance videos. The model first extracted the representative aggressive behavior from the surveillance video by I3D feature extraction network, and then input the features into the temporal pyramid network to obtain multi-scale temporal information, and finally used coarse prediction to obtain the initial nomination, and used fine prediction to refine the obtained coarse nomination, and the coarse prediction regressed the frame position of the action interval, the offset from the start and end of the action, and the occurrence of the action by the temporal convolution network. The fine prediction refined the boundary positions and obtained the final prediction results by activation-guided learning and boundary contrast learning. In order to train and validate the proposed model, a video dataset containing 174 videos of different durations and 464 temporal annotations for the detection of aggressive behavior among gregarious pigs was constructed. The experimental results showed that the model could achieve a recall rate of 79.1% at average tIoU when the number of nominations was 100, and detected 90 min of original surveillance video containing 10 segments of aggressive behaviors, and the prediction results could cover all real instances, which could better detect the aggressive behavior among gregarious pigs. This study can provide a reference for modern pig farms to achieve intelligent analysis and healthy breeding.

Key words: gregarious pigs, aggressive behavior, video recognition, behavior detection