基于Cow-DETR与深度图像的非接触式奶牛体质量评估

Non-contact Predicting Method of Dairy Cow Weight Based on Cow-DETR and Deep Image

针对当前牧场奶牛体质量(体重)称量效率低,人工参与容易引发奶牛应激等问题,提出了一种基于改进DETR(Detection transformer)网络的端到端式奶牛体质量评估方法(Cow-DETR),实现利用奶牛背部深度图像进行非接触式奶牛体质量评估。首先设计并搭建实验数据采集装置,利用Intel RealSense D435深度相机和体重秤采集奶牛背部深度图像和体质量数据;然后,通过边缘平滑滤波器和孔洞填充滤波器对深度图像进行补全处理,减少深度数据缺失对体质量评估的影响;最后,以DETR网络为基础建立奶牛体质量评估模型,通过在预测模块中添加含有交替全连接层的体质量预测单元,提升奶牛体质量相关的特征信息提取能力,实现端到端式奶牛背部定位的同时进行奶牛体质量非接触式评估。结果表明,本文方法可以实现较高精度的奶牛体质量评估,通过5倍交叉验证,在含有139头奶牛数据的数据集中,平均绝对误差不超过17.21 kg,平均相对误差不超过3.71%,单幅图像平均识别时间为0.026 s。通过与现有体质量评估方法相对比,本文方法比其他6种方法在更多的奶牛头数的数据集中取得了更低的平均绝对误差和平均相对误差,同时本文方法对奶牛站立姿势要求较低,更符合牧场实际生产需要,为奶牛体质量评估提供了新的解决思路。

In order to solve the existence of some problems such as the low weighing efficiency of dairy cows in the current pasture,and being easy to cause the stress of dairy cows by manual participation,an end-to-end method of dairy cow weight estimation(Cow-DETR) based on improved detection transformer(DETR) network was proposed.The non-contact estimation on the dairy cow weight was carried out by using the depth image of dairy cow back.Firstly,a data acquisition device was designed and built,with which the cow's back depth image and weight data were collected by using the Intel RealSense D435 depth camera and the weight scale.Then,deep image data was filled by using the edge flat filter and hole filling filter to reduce the impact of deep data loss on weight estimation.Finally,by adding the weight prediction unit with an alternate fully connection layer(AFC) to the prediction module of DETR to establish a cow weight estimation model.AFC was added to improve the ability of dairy cow weight-related feature extracting.It implemented the end-to-end dairy cow back positioning while performing a non-contact estimation of dairy cow weight by Cow-DETR model.The data of 139 cows were used to evaluate the model,and the results through 5-fold cross validation showed that the weight estimation method proposed can achieve a high accuracy in dairy cow weight estimation.The average absolute error of weight estimation was below 17.21 kg,the average relative error was less than 3.71%.The average recognition time was 0.026 s per image.Compared with the existing weight estimation methods,the results showed that Cow-DETR got lower average absolute error and average relative error than the other six methods in more dairy cow data.In the meantime,the method proposed had less requirements on the posture of dairy cow,which can more comply with the actual production demand of ranch and provide a solution for the weight estimation of dairy cow.