基于深度学习YOLOv3的奶牛犊行为检测

Deep learning YOLOv3-based behavior detection for dairy cows

为了实时监测奶牛犊健康状况,试验采用萤石摄像头对散养的6月龄左右的5头奶牛犊视频进行采集,将标注的奶牛犊行为图像进行训练及验证,建立基于YOLOv3和Darknet53网络的奶牛犊行为识别模型。在5头奶牛犊视频中随机选择4409幅奶牛犊图像进行躺卧、挤头、吃食、饮水和站立行为标注,并按9∶1划分训练集和验证集。其中3969幅用作训练集,共计19845个奶牛犊行为框,其余440幅用作验证集,另取108幅无标注的奶牛犊图像作为测试集。结果表明:采用YOLOv3进行奶牛犊行为识别和python语言编程进行模型训练时,训练阈值为0.2时,在测试集上目标检测的平均准确率为85.6%。说明奶牛犊行为识别模型的实时性较好,能够识别奶牛犊的不同行为,有助于及时了解奶牛犊的健康状况,实现奶牛犊的精准养殖。

In order to monitor the health status of calves in real time,fluorite cameras were used to collect videos of 5 calves aged about 6 months in free range breeding conditions.Training and verification of the marked little cow behavior images were conducted to establish a calf behavior recognition model based on YOLOv3 and Darknet53 networks.From the videos of 5 calves,4409 images were randomly selected to mark the behavior of lying,squeezing,eating,drinking and standing,and the training set and verification set were divided according to 9∶1.Of these,3969 were used as training sets,with a total of 19845 calf behavior frames.The remaining 440 were used as validation sets,and another 108 images of unmarked calves were taken as test sets.The results showed that average accuracy rate of target detection on test set is 85.6%when YOLOv3 was used for behavior recognition and python language programming was used for model training,and the training threshold was 0.2.The results indicated that the behavior recognition model had a good real-time performance and could recognize different behaviors of the little cows,which is conducive to timely understanding the health status of the little cows and realizing accurate breeding of the little cows.