利用Kalman滤波和Hungarian算法的多目标奶牛嘴部跟踪及反刍监测

针对奶牛养殖场复杂环境下多目标奶牛嘴部自动跟踪及反刍监测的困难，该研究提出了一种基于嘴部区域跟踪的多目标奶牛反刍行为智能监测方法。在YOLOv4模型识别奶牛嘴部上下颚区域的基础上，以Kalman滤波和Hungarian算法跟踪上颚区域，并对同一奶牛目标的上颚和下颚区域进行关联匹配获取嘴部咀嚼曲线，以此获取反刍相关信息，从而实现多目标奶牛个体的嘴部跟踪和反刍行为监测；为解决奶牛快速摆头运动和棚舍栏杆遮挡引发奶牛标号变化的问题，提出未匹配跟踪框保持及扩大的方法。采集并选择实际养殖场环境下的反刍奶牛视频66段，对其中58段视频采取分帧操作得到图像，制作YOLOv4模型数据集，以其余8段视频验证跟踪方法和反刍行为判定方法的有效性。试验结果表明，YOLOv4模型对奶牛嘴部上颚、下颚区域的识别准确率分别为93.92%和92.46%；改进的跟踪算法可实现复杂环境下多目标奶牛嘴部区域的稳定跟踪，且有效解决了栏杆遮挡、快速摆头运动造成的奶牛标号变化现象，上下颚匹配率平均为99.89%，跟踪速度平均为31.85帧/s；由反刍行为判定方法获取的咀嚼次数正确率的平均值为96.93%，反刍时长误差的平均值为1.48 s。该研究可为实际养殖中多目标奶牛反刍行为的智能监测和分析提供参考，也可供其他群体动物运动部位的跟踪和行为监测借鉴。

Multi-target cow mouth tracking and rumination monitoring using Kalman filter and Hungarian algorithm

Rumination enables cows to chew grass more completely for better digestion, thereby closely relating to the health, production, reproduction, and welfare of cows. To perceive ruminant behavior has widely been one of the most important steps in modern dairy farm management. However, the traditional monitoring ruminant behavior of cows depends mainly on human labor, time-consuming and laborious. In this study, feasible intelligent monitoring was proposed for multi-target automatic tracking mouth and ruminant behavior of cows in the complex environment of dairy farms using the Kalman filter and Hungarian algorithm. The upper and lower jaw regions of cow mouths were firstly recognized by the YOLOv4 model. Subsequently, the region of the upper jaw was tracked by the Kalman filter and Hungarian algorithm. The chewing curve of the mouth region was then obtained to match the upper and lower jaw regions of the same cow. Finally, the related rumination information was achieved to further realize the mouth tracking and ruminant behavior monitoring of multi-target cows. As such, the unmatched tracking boxes were remained and expanded to deal with the identity change of cows caused by rapid head swing or shed railing occlusion. 66 videos were collected for ruminant cows in the actual farm environment, where 58 videos were divided into frames to make the dataset for the YOLOv4 model, and the remaining 8 videos were used to verify the tracking and rumination behavior. Videos data was involved sunny, cloudy, and rainy days, in which the number of cows varied from 2 to 3. Lying or standing was included in the ruminant posture of cows. Additionally, there were some interference factors, such as the rapid head swing of ruminating cows, shed railing occlusion, and the movement of other cows. Two indexes were selected to evaluate the detection performance of the YOLOv4 model, including average precision and mean average precision. 6 400 images of the dataset were trained, and 800 images were tested. The results showed that the average precisions of the YOLOv4 model were 93.92% and 92.46% for the detection of the upper and lower jaw region, respectively. The mean average precision of YOLOv4 reached 93.19%, which was 1.04, 4.25, and 1.74 percentage points higher than that of YOLOv5, SSD, and Faster RCNN models, respectively. Four indexes were selected to verify the performance of tracking and rumination behavior under different environments, including the rate of identity switch, the rate of identity match, the detection rate of chewing times, and the tracking speed. It was found that the YOLOv4 model realized the stable multi-target tracking of mouth regions of cows in complex environments, while effectively alleviating the identity change of cows resulted from the rapid head swing and shed railing occlusion. The average rate of identity match was 99.89% for the upper and lower jaws, and the average tracking speed was 31.85 frames/s. The average detection rate of chewing times was 96.93% after the evaluation of rumination behavior, and the average error of rumination time was 1.48 s. This finding can provide a strong reference for the intelligent monitoring and ruminant behavior of multi-target cows (or moving parts of animals) in actual breeding.