基于无锚时序动作定位的群养生猪争斗行为检测研究

Research on detection of aggressive behavior among gregarious pigs based on anchor-free temporal localization

现代生猪养殖业不断趋于规模化和集约化,密集型群养环境下生猪的争斗行为频繁发生,严重影响生猪健康、福利、生产性能以及养殖场经济效益。以群养猪舍监控视频为研究对象,提出一种基于端到端的无锚时序动作定位框架的群养生猪争斗行为检测模型,以实现从群养生猪监控视频中自动检测出争斗行为的发生及其发生时段。该模型首先通过I3D特征提取网络提取监控视频中具有代表性的争斗行为特征,随后将该特征输入时序金字塔网络获取多尺度的时序信息,最后使用粗糙预测获取初始提名,使用精细预测对得到的粗糙提名进行细化,粗糙预测通过时序卷积网络回归出动作区间的帧位置、距离动作开始和结束位置的偏移量以及所发生的动作类别,精细预测通过激活指导学习和边界对比学习来对边界位置进行细化并得到最后的预测结果。为了训练和验证所提出模型,构建包含174段不同时长视频和464段时序标注的群养生猪争斗行为检测视频数据集。试验结果表明,该模型能够在提名数量为100段时在平均tIoU下达到79.1%的召回率,对包含10段争斗行为的90 min的原始监控视频进行检测,预测结果能覆盖所有真实实例,能较好地完成群养生猪争斗行为检测的任务。本研究可为现代化生猪养殖场实现生猪行为智能分析与健康养殖提供参考和借鉴。

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.