Sheep Pose Estimation Based On Cascaded Neural Network

With the wide application of deep learning in pose estimation,the research object can be no longer limited to human,the method of human pose estimation can be extended to sheep.Sheep pose estimation is a process of detecting keypoints from the input image and connecting them by line segments.China is a big country in agriculture and animal husbandry.Compared with traditional breeding methods,intelligent breeding can improve animal health and increase breeding efficiency.According to the pose of sheep,we can infer the movement time,standing and lying times of sheep,and then we can infer their health and living conditions.At present,lack of relevant research on sheep pose estimation,and there is no available dataset for sheep pose estimation.At the same time,the sheep’s living environment is complex and changeable,and the keypoints of the body are occluded and similar to the environment,which affects the detection accuracy of the keypoints on sheep.In this thesis,by taking the life video of sheep on the spot,we make the dataset of sheep pose estimation and improve the human pose estimation method to realize the sheep pose estimation.The research contents and conclusions are as follows:1)Make Sheep pose estimation dataset.After clipping the video,select the complete sheep image without large area occlusion of keypoints as the original material of the dataset.According to the keypoints selection scheme of human pose estimation,use Landmark＿Annotation to annotate sheep pose keypoints estimation dataset.2)The model of sheep pose estimation based on cascaded residual network.The analysis of Res Net-50 and Re Fine Net in the existing human pose estimation methods is carried out to extract and refine the keypoints of the sheep pose estimation.The original online hard keypoints mining(OHKM)is improved.In the second stage,only the keypoints in the pre ordered network are trained to improve the average detection accuracy of keypoints in the network model.3)The model of sheep pose estimation model based on cascaded hourglass network.Referring to the idea of human pose estimation,due to the great difference of detection accuracy of keypoints in sheep pose estimation,the online hark keypoints mining is added into the model in the second stage of training.In the first stage,the parts with good detection accuracy of key points are no longer back propagation to update parameters,so as to train the key points with poor detection effect.In this thesis,4200 sheep images and keypoints annotation files are produced for the research and application.In order to verify the effect of the improved model on the sheep pose estimation dataset,the comparison experiment is set on the original model and the improved model on the sheep pose estimation dataset.The experimental results show that the improved model in Chapter 4 and Chapter 5 can reach the average detection accuracy of 71.2% and 70.3% in the dataset;In addition,two improved models and mainstream human pose estimation models were set to verify the effect of the model.The experimental results show that the average detection accuracy of the improved sheep key points is better than the comparison models.