基于深度学习的群养鸡只行为监测方法研究

Research of behavior monitoring method of flock hens based on deep learning

畜禽的行为能够反映其健康状况、环境舒适度等福利信息,是评价畜禽福利状况的重要指标。为了实时自动监测鸡舍中鸡群的行为状况,提出了1种基于深度学习的鸡群行为监测方法。用摄像头连续4个月每天7:00—17:30记录散养鸡群在鸡舍内的活动,筛选7988幅图片对鸡只的采食、站立、趴卧、梳羽、啄羽和打架行为进行标注,利用YOLO v4目标检测模型识别,各行为平均精确率分别为采食96.67%、站立90.34%、趴卧78.46%、梳羽82.01%、啄羽63.38%、打架67.14%,模型总体平均精确率为79.69%。为进一步提高啄羽行为检测结果,采用了时间序列分析方法提取持续时间大于30 s的啄羽行为。实验结果表明,该方法可以实现群养鸡只行为的实时自动监测,解决人工不能全天实时监测的问题,为精准化养殖提供了可能。

The behavior of livestock and poultry can reflect their health condition, environmental comfort level and other welfare information, which is an important index to evaluate the welfare status. In order to monitor the behavior of the chickens in the chicken house automatically and in real time, a deep learning method was proposed to monitor the behavior of chickens. The camera recorded chicken activities everyday from 7:00 to 17:30 for four months in the chicken house. Eating, standing, lying, preening, feather pecking and fighting behaviors of chickens were labeled in 7988 images, the average precision(AP) rates were 96.67%, 90.34%, 78.46%, 82.01%, 67.14% and 63.38% respectively, and the mean average precision(mAP) was 79.69%. To improve the feather pecking behavior detection, a method of time series analysis was used, which could extract the feather pecking behavior lasting more than 30 s. The experimental results show that this method can realize real-time automatic monitoring of the behavior of chickens, and solve the problem that human beings cannot monitor the behavior of chickens all day long, which provides the possibility for precision livestock farming.