Multi-target Tracking Of Pigs Under Occlusion

With the rapid development of deep learning,multi-target tracking technology is also gradually applied to major industries.It is widely used in intelligent breeding,human-computer interaction,security monitoring and other fields.In recent years,with the continuous improvement of people’s living needs,the scale of pig breeding of enterprises has also expanded rapidly.The traditional breeding method has not met the needs of management.Enterprises are also looking for an efficient and intelligent breeding method.Therefore,this paper applies the multi-target tracking technology to the field of pig breeding,and designs a multi-target tracking network for pigs.It effectively reduces the number of contact between pig farmers and pigs,and prevents human infection to pigs.The main work is as follows:The horizontal truth box pig data set was made.The pig multi-target tracking algorithm based on Faster RCNN used this data set for training and testing.Then,according to the test results,a detailed analysis was carried out to find the cause of the problem.First,too many impurities were leaded into the horizontal truth box in the process of making the data set;Second,NMS will inhibit pigs with serious adhesion of horizontal detection box.By reading the literature in the field of remote sensing image detection,it was found that the advantages of rotating target detector can perfectly solve the problems in pig multi-target tracking algorithm based on Faster RCNN.A new algorithm for pig multi-target tracking was redesigned.After that,the pig data set of multi-angle truth box was made for the training and testing of R3 Det.According to the movement characteristics of pigs,the box regression network was redesigned.Re-identification network was redesigned according to the appearance characteristics of pigs.According to the characteristics of rotating target detector,the NMS of the original network were replaced.Finally,the experimental results show that the pig multi-target tracking algorithm designed in this paper was better than the pig multi-target tracking algorithm based on Faster RCNN,MOTA and MOTP are as high as93.3% and 91.9% respectively,and the ID switches are reduced to 1.