基于改进YOLO v7-Pose的牛脸关键点检测与姿态识别方法

Cow Face Keypoint Detection and Pose Recognition Based on Improved YOLO v7-Pose

摘要:

奶牛脸部关键点检测在牛场智能化中发挥着重要的作用，它可以帮助实现牛脸识别、牛脸对齐、头部动作检测与行为识别等。针对目前奶牛养殖环境中存在牛脸遮挡、弱光照等问题，提出了一种改进的YOLO v7-Pose网络模型的算法，可用于牛脸关键点检测和头部姿态识别。首先通过网络相机在牛场获取奶牛脸部图像并构建数据集。其次，在YOLO v7-Pose网络模型中引入SPPFCSPCL结构，以提高奶牛脸部关键点的特征提取能力；将关键点检测的损失函数OKS替换为WingLoss损失函数，增加奶牛脸部关键点的检测精度；最后，使用L1范数对改进的模型剪枝，使改进后的网络模型参数量降低。试验结果表明：改进模型YOLO v7-SCLWL-Pose检测牛脸关键点较原模型AP提升5个百分点，AP0.5提升2.7个百分点，改进后模型内存占用量仅为106.7MB，减少33.6%。将本文关键点检测用于姿态识别，试验结果对抬头和低头等动作的识别准确率达到95.5%和86.5%。本研究为牧场奶牛行为识别提供了支撑技术。

Abstract:

Facial keypoint detection in dairy cows plays a crucial role in the automation of cow farms. It aids in cow face recognition, face alignment, head movement detection, and behavior recognition. In view of the problems of cow face occlusion and weak light in the current dairy farming environment, an improved algorithm of YOLO v7-Pose network model was proposed, which can be used for keypoint detection and head pose recognition of cow face. Firstly, dairy cow facial images were collected from cow farms by using network cameras and a dataset was constructed. Secondly, the SPPFCSPCL structure was integrated into the YOLO v7-Pose network model to enhance its feature extraction capabilities for cow facial keypoints. The WingLoss loss function replaced the OKS loss function for keypoint detection, thereby improving the accuracy of cow facial keypoint detection. Finally, L1 regularization was applied to prune the improved model, reducing the number of network parameters. The experimental results showed that the cow face keypoint detection of improved model YOLO v7-SCLWL-Pose was improved by 5 percentage points and AP0.5 was improved by 2.7 percentage points compared with the original model AP, and the memory occupation of the improved model was only 106.7MB, which was reduced by 33.6%. The keypoint detection was applied to pose recognition, and the experimental results showed that the recognition accuracy of the motions of looking up and looking down reached 95.5% and 86.5%. This research can provide support technology for behavior recognition in dairy cows on farms.