Pig Multi-objective Based On YOLOv4DeepSORT Research On Detection And Tracking Algorithm

With the popularization of smart agriculture,intelligent farming has been gotten more and more attention.It is of great significance to detect the abnormal behaviors of pigs quickly and accurately,so as to warn pigs of diseases and take corresponding measures in time.In this thesis,deep learning technology was used to study the multi-target detection and tracking method of herd-farmed pigs in the actual feeding environment.The main research contents are as follows:1)Mini＿YOLOv4 method was proposed for daily behavior detection of pigs.The Mini＿YOLOv4 algorithm proposed in this thesis firstly used the lightweight Mobile Netv3 network structure to replace the feature extraction network of the original YOLOv4 to reduce the overall number of model parameters;then,in the detection of CBL＿block1 and CBL＿block2 modules of network,the depth-separable convolution was used to replace the traditional convolutional operation to reduce the insufficient memory and high delay caused by the complex model.Finally,the last layer of 3×3 convolution in each detection scale of the original YOLOv4 network was replaced by the Inception structure to improve the accuracy of the model in pig posture detection.The experimental results showed that the proposed Mini＿YOLOv4 improved the detection accuracy by 3.83 percentage points compared with the original YOLOv4 algorithm.The detection speed of improved model was 32 frames/s higher than that of the original YOLOv4 algorithm.2)A real-time pig multi-target tracking algorithm based on DeepSORT was proposed.The accurate and fast Mini＿YOLOv4 model was used to replace the original Faster R-CNN target detection algorithm as the detector in DeepSORT algorithm,and CIOU was used in the cascaded matching module to improve the IOU in the original algorithm to cope with the problems of shielding and missing detection in the process of pig target tracking.Kalman filtering and depth feature matching were used to complete the association of pig target frames in the actual feeding environment,the average tracking accuracy was 69.3% in the four test video segments with different numbers of targets as well as different levels of complexity,and the jump of target ID could be effectively suppressed,indicating that it was feasible to apply the DeepSORT algorithm to the efficient tracking of pig targets in the actual feeding environment.3)Implementation of multi-target detection and tracking system for pigs.Based on the results of posture detection and multi-objective tracking of pigs,the daily resting time,feeding and drinking frequency of pigs were statistically analyzed,and the daily resting time,feeding and drinking frequency evaluation table and health status score table of pigs were established by integrating the relevant literature data through the behavior anchoring level evaluation method.Finally,Py Qt5 was used to design the user page of the system,which contains three modules of pig posture detection,target tracking and health status evaluation.