Research On Beef Cattle Skeleton Extraction Method Based On Multi Target Tracking

The behavior of beef cattle can reflect their physical condition and health level.How to obtain the behavior of beef cattle has become a research hotspot in recent years.Skeleton information extraction can be used as the basis of behavior recognition.At the same time,it has the characteristics of high accuracy and high efficiency to obtain beef skeleton information through monitoring equipment.Therefore,it is of great significance to obtain the skeleton information of beef cattle accurately and quickl y by using monitoring equipment for the development of beef cattle breeding industry.In this study,the beef cattle skeleton extraction based on multi-target tracking was carried out by using the farm monitoring video,which provided the basis for identifying the behavior of beef cattle in the farm through video monitoring.The main work of this thesis is as follows:(1)Design of beef cattle multi-target tracking method based on Deep SORT algorithm.In order to solve the problem of poor tracking effect caused by uneven distribution and large scale change of beef cattle,a long short range context enhancement module(LSRCEM)is proposed to improve the YOLOv3 algorithm,expand the receptive field of the model,and obtain the LSRCEM-YOLO target detection algorithm,which increases the detection accuracy of the model,Then,we use Mobile Net V2 to replace Dark Net53 as the backbone network,which reduces the complexity of the model.Combined with Mudeep re-recognition model,we achieve multi-target tracking of beef cattle.The results showed that the map value of LSRCEM-YOLO was 92.3%,and the model parameters were only 10%of YOLOv3,which was 31.34% lower than YOLOv3-tiny;Compared with the Deep SORT algorithm,the MOTA index of multi-target tracking is improved from 32.3% to 45.2%,and the number of ID switch is reduced by 69.2%.(2)Design of beef cattle skeleton extraction method based on Alpha Pose algorithm.Aiming at the problem of low accuracy of beef skeleton extraction,firstly,LSRCEM-YOLO algorithm is used to detect individual beef cattle and extract a single target.Then,16 key points of beef cattle are defined to establish a beef skeleton extraction data set.The extracted single beef target is input into stacked hourglass module(SHM)for skeleton extraction,Finally,all the single cattle skeleton extraction results are mapped to the beef cattle multi-target tracking video to realize the beef cattle skeleton extraction based on multi-target tracking.The experimental results show that the skeleton extraction models based on deep learning stack hourglass network and Alpha Pose model achieve 90.39% and77.33% accuracy in single cow and multi cow skeleton extraction tasks respectively.The beef cattle skeleton extraction method based on multi-target tracking constructed in this study can provide new ideas for intelligent breeding of beef cattle,and provide effective support for accurate feeding and intelligent breeding of beef cattle.