基于视频分析的多目标奶牛反刍行为监测

奶牛反刍行为与其生产、繁殖和应激行为等存在较强的相关性,现有方法多采用人工观察或可穿戴式装置进行奶牛反刍行为的监测,存在误差大、容易引起奶牛应激反应、成本高等问题。为了实现多目标奶牛反刍行为的实时监测,该研究基于视频分析与目标跟踪技术,在获取奶牛嘴部区域的基础上,分析对比了压缩跟踪算法(compressive tracking,CT)和核相关滤波算法(kernelized correlation filters,KCF)在多目标奶牛反刍监测中的性能。为了验证不同算法对奶牛反刍行为监测的效果,分别用9段视频进行了试验,针对误检问题提出了有效的咀嚼次数判定模型,最后与实际的奶牛反刍数据进行了对比。试验结果表明:对多目标监测,KCF算法平均帧处理速度为7.37帧/s,是CT算法平均帧处理速度0.51帧/s的14.45倍;KCF算法平均误差为13.27像素,是CT算法平均误差38.28像素的34.67%。对双目标监测,KCF算法的平均误检率为7.72%,比CT算法的平均误检率18.56%低10.84个百分点;2种算法的帧处理速度分别为10.11帧/s和0.87帧/s;平均跟踪误差分别为22.19像素和28.51像素,KCF算法的平均跟踪误差仅为CT算法的77.83%。试验结果表明,KCF算法具有较低的误检率及较高的帧处理速度,更适合奶牛反刍行为的监测。在此基础上,验证了2种算法在不同光照、不同姿态和不同程度遮挡等影响因素下的监测效果,结果表明,CT算法会出现不同程度的偏离,甚至丢失目标,而KCF算法仍然具有良好的效果和较好的适应性,表明将KCF算法应用于全天候多目标奶牛反刍行为的分析是可行的、有效的。

Monitoring of multi-target cow ruminant behavior based on video analysis technology

Abstract: There is a strong correlation between the ruminant behavior of dairy cows and their production, reproduction, estrus and stress behaviors. Rumination rhythm and time are related to the health status of dairy cows closely. By collecting and analyzing ruminant signals, it is possible to detect the physiological activities of dairy cows accurately and find the health problems of dairy cows in time. It is of great significance to improve the modern management level of dairy cows, promote the fine management of dairy cows'' breeding, and improve the efficiency of pastures. The existing methods mostly use artificial observation or wearable devices to monitor the ruminant behavior of dairy cows, which has the problems of large error, being easy to cause stress reaction of dairy cows, high cost, low real-time performance, and so on. In the field of target recognition and tracking, the kernelized correlation filters (KCF) algorithm and the compressive tracking (CT) algorithm are widely used and have achieved good results, such as high real-time performance, high accuracy, effective suppression of tracking drift, high robustness, and good tracking effect. In order to achieve real-time multi-target monitoring of ruminant behaviors of dairy cows, by video analysis and target tracking technology, on the basis of obtaining the mouth area of dairy cows, the performances of CT algorithm and KCF algorithm in multi-target intelligent monitoring of cows ruminating were analyzed and compared in this study. To verify the effect of different algorithms on the monitoring of ruminant behavior of dairy cows, 9 videos were used to test and then compared with the actual ruminant data of cows, including 2 multi-target cow videos and 7 double-target cow videos. Additionally, aimed to the occurrence of missed detection, false detection, and so on, we proposed an effective judgment model for counting the number of chewing times. The test results showed that for multi-target monitoring, the average frame processing speed was 7.37 frames/s with the KCF algorithm, and was 0.51 frames/s with the CT algorithm; the average error of the KCF algorithm was 13.27 pixels, and that of the CT algorithm was 38.28 pixels; the average tracking error of the KCF algorithm was 34.67% of that of the CT algorithm. For double-target monitoring, the maximum false detection rate of the KCF algorithm was 18.42%, and the lowest was 0; the highest false detection rate of the CT algorithm was 81.58%, and the lowest was 0; the average false detection rate of the KCF algorithm was 7.72%, which was 10.84 percentage points lower than that of the CT algorithm. The frame processing speeds of the 2 algorithms were respectively 10.11 and 0.87 frames/s; the highest tracking errors were 45.80 and 46.13 pixels respectively, and the lowest were 7.71 and 17.33 pixels respectively. The average tracking error of the KCF algorithm was only 77.83% of that of the CT algorithm. The experimental results showed that for multi-target monitoring of rumination behavior of cows in a complex environment with the requirements of high accuracy and high real-time performance, the KCF algorithm with the low false detection rate and high frame processing speed was more suitable. On this basis, we verified the effectiveness of these 2 algorithms in monitoring ruminant behavior when cows were exposed to different lighting, and had different postures and different degrees of occlusion. The results showed that the CT algorithm had different degrees of deviation, and even lost the target, while the KCF algorithm still had good results and good adaptability in the nighttime video tracking. It shows that it is feasible and effective to apply the KCF algorithm to the all-day multi-target analysis of the ruminant behavior of dairy cows.