基于面部关键点轨迹的鱼骨式挤奶厅奶牛异常行为检测方法

Method for Detecting Abnormal Cow Behavior in Fishbone-style Milking Parlors Based on Facial Landmark Trajectories

摘要:

挤奶过程是奶牛应激反应和动物福利问题的高发环节,对异常行为进行自动化监测对于提升挤奶管理水平和保障奶牛福利具有重要意义。 针对鱼骨式挤奶厅中空间受限、多牛同时入镜、遮挡频繁以及个体轨迹易跳变等问题,本文面向鱼骨式挤奶厅特定场景,开展了基于面部关键点轨迹的奶牛异常行为检测研究。 以 DeepLabCut 作为关键点感知模块,在此基础上引入栏位位置约束与牛脸朝向约束进行多牛轨迹修正,并结合场景化行为特征与规则判别实现事件级异常输出。 实验结果表明,基于栏位位置和牛脸朝向的双约束融合修正可将多奶牛编号跳变次数由 18. 500 次降至 0. 125 次,准确率由 93. 83% 提升至 99. 96% ;在异常行为事件级检测中,单牛异常行为事件级的精确率、召回率、漏检率和误检率分别为 96. 55% 、96. 55% 、3. 45% 和 3. 45% ;多牛交互异常行为事件级的精确率、召回率、漏检率和误检率分别为 85. 71% 、100% 、0 和 14. 29% ;关键点感知模块的精确率、召回率、平均精度均值和标准化平均误差分别达到 98. 69% 、98. 61% 、97. 22% 和 5. 31% 。 该方法为固定视角挤奶场景下的异常行为自动检测提供了一种可行方案,具有较好的可解释性和工程适用性。

Abstract:

The milking process is a high-risk phase for stress responses and animal welfare issues in dairy cows; automated monitoring of abnormal behaviors is crucial for improving milking management and ensuring cow welfare. Aiming to address the challenges in herringbone milking parlors, such as limited space, multiple cows entering the field of view simultaneously, frequent occlusions, and erratic individual trajectories, the detection of abnormal cow behaviors was investigated based on facial landmark trajectories, specifically tailored to the unique environment of herringbone milking parlors. Using DeepLabCut as the keypoint detection module, stall position constraints and cow-face orientation constraints were introduced to correct multi-cow trajectories, and combined scenario-specific behavioral features with rule-based classification to achieve event-level anomaly detection. Experimental results showed that the dual-constraint fusion correction based on stall position and cow face orientation reduced the number of keypoint ID changes for multiple cows from 18. 500 to 0. 125, while increasing accuracy from 93. 83% to 99. 96% . In event-level detection of abnormal behavior, the precision, recall, false negative rate, and false positive rate for single-cow abnormal behavior events were 96. 55% , 96. 55% , 3. 45% , and 3. 45% , respectively. For multi-cow interactive abnormal behavior event-level detection, the precision, recall, false negative rate, and false positive rate were 85. 71% , 100% , 0, and 14. 29% , respectively. The keypoint detection module achieved a precision of 98. 69% , a recall of 98. 61% , a mean average precision of 97. 22% , and a standardized mean error of 5. 31% . This method can provide a viable solution for the automatic detection of abnormal behaviors in fixed-view milking scenarios, offering good interpretability and engineering applicability.