基于PointNet++和改进ConvNeXt模型的奶牛个体身份识别方法

Individual Cow Identification Method Based on PointNet++ and Improved ConvNeXt Network

摘要:

奶牛个体身份识别是实现精准智能养殖的前提和基础，但基于图像信息的身份识别方法容易受环境和拍摄角度的影响。为了实现顶视条件下奶牛身份的精准识别，提出了一种基于PointNet++和改进ConvNeXt模型的奶牛个体身份识别方法。首先，采集奶牛顶视RGBD图像，使用PointNet++模型对奶牛的腰角骨点和臀角骨点进行粗定位；然后，分析腰角骨和臀角骨区域的曲率变化精准定位腰角骨点和臀角骨点，根据腰角骨点和臀角骨点距离关系确定关键区域，将关键区域转为二维体斑图像；最后，基于改进ConvNeXt模型进行图像分类，实现奶牛身份的精准识别。对30头奶牛的总共6800幅顶视图像构建数据集，以比例7∶2∶1构建训练集、验证集和测试集。结果表明，点云分割模型平均精确率AP50为92.5%，奶牛身份识别准确率可达94.67%，改进ConvNeXt模型与原始模型相比，在权重基本一致的情况下，分类精度提高4.83个百分点。该方法对顶视视野中奶牛位置和角度具有较高的鲁棒性。

Abstract:

The identification of individual cows is a prerequisite and foundation for realizing accurate and intelligent farming, but the identification method based on image information is easy to be affected by the environment and observation angle. In order to achieve accurate identification of cow identity under top-view conditions, an individual identification method of cow based on PointNet++ and improved ConvNeXt model was proposed. Firstly, the apex RGBD images of cows were collected, and PointNet++ model was used to locate the hook and pin bones of cows. Secondly, the curvature changes of hook and pin were analyzed to accurately locate hook and pin, and the key areas were determined according to the distance relationship between hook and pin, and the key areas were converted into two-dimensional body spot images. Finally, based on the improved ConvNeXt model, image classification was performed to achieve accurate identity recognition. A total of 6800 top view images from 30 cows were constructed, and the training set, validation set, and test set were constructed at a ratio of 7∶2∶1. The results showed that the AP50 of the point cloud segmentation model was 92.5%, and the identification accuracy of the cow can reach 94.67%. Compared with that of the original model, the classification accuracy of the improved ConvNeXt model was improved by 4.83 percentage points under the condition that the weight was basically the same. The method had high robustness to the position and angle of the cow in the top visual field.