| China is a big country for raising pigs,and the scale of pig breeding and pork production are among the top in the world.The health status of pigs is an important factor affecting the production and quality of pork,and the feeding behavior of pigs is an important basis for evaluating the health level of pigs.Therefore,in the process of pig breeding management,monitoring the health level of pigs through the feeding behavior of pigs is particularly important.important.At present,the health monitoring of pigs is mainly based on manual inspection observation and application of contact sensor monitoring.Manual monitoring is time-consuming and labor-intensive,and is highly subjective and prone to misjudgment.The application of contact sensor monitoring has certain damage to animal welfare,and it is easy to be caused by Changes in the external environment produce noisy data.In view of the above problems,based on the computer vision platform,this paper uses deep learning technology to carry out research on multi-target tracking methods for group-raising pigs,research on groupraising pig head recognition methods,and group-raising pig feeding behavior recognition methods.Feeding behavior information provides a new idea to improve the health monitor level of group pigs.The main research contents and conclusions of this paper are as follows:(1)Research on multi-target tracking method for group pigs.In view of the current multitarget tracking of group pigs,most of which are based on Anchor Box identification algorithms,the model is complex,the results are unstable,and there is a deviation between the actual target center and the Anchor Box center in the process of multi-target tracking.A multi-target tracking algorithm that uses an Anchor Free-based target detection method to determine the target center of the input image.The results show that the MOTA(multi-target tracking accuracy)of the FairMot multi-target tracking algorithm is 59.3%,the MOTP(multi-target tracking accuracy)is 74.7%,and the IDSwitch(target ID hopping times)is 167 times.For the Deepsort multi-target tracking algorithm of YOLOv4,its MOTA and MOTP are increased by10.7% and 18.9% respectively,and IDSwitch is decreased by 34.5%.It shows that the FairMot multi-target tracking algorithm can effectively suppress the target ID jump,and the FairMot multi-target tracking algorithm is feasible for multi-target tracking of group pigs,which lays a foundation for group pig feeding behavior recognition.(2)Research on head recognition method of group pigs.Aiming at the problem of misidentification caused by being easily occluded by other pigs in the process of group pig head recognition,this study uses VGG16 convolutional neural network to establish a group pig head recognition model.Firstly,the image of the target bounding box is intercepted according to the trajectory coordinates obtained by the multi-target tracking of the group pig,and the image of the target bounding box of the group pig is intercepted by Open CV to obtain the patch image of the four corners of the target bounding box of each pig and input to the VGG16 network for grouping.Pig head identification.The accuracy rate of VGG16 neural network model recognition is 95.91%,the recall rate is 92.54%,and the running speed is 58.62%higher than that of the classic group pig head recognition method based on Io U calculation,indicating that the model is suitable for grouping in structured scenarios.Accurate identification of pig heads.(3)An algorithm for identifying feeding behavior of non-contact group pigs based on pig head occupancy index was proposed.In view of the current shortage of feeding behavior monitoring of group pigs,this paper studies the problem of non-contact group pig feeding behavior recognition based on video analysis,and designs a feeding behavior recognition algorithm based on pig head occupancy index.In the structured pig house environment,the feeding behavior of group-raised pigs only occurs in the feeding trough area of the pig house.When the pig head is located in the feeding trough area,the feeding behavior is determined by the pig head occupancy index.The results show that the recognition accuracy of the algorithm on the multi-segment video test set is 100%,and the recall rate is 96.27%.On the test set of all frame sequences of a single video,the average recognition accuracy of pig individual targets is 98.95%,and the average recall rate is 81.52%.The identification results of the algorithm are consistent with the manual observation results,indicating that the algorithm is suitable for accurate identification of feeding behavior of group pigs in structured scenarios.(4)A feeding behavior recognition system for group-raised pigs in a structured environment was designed.Using the method proposed in this paper to identify the feeding behavior of group pigs,a system for identifying feeding behavior of group pigs in a structured environment is designed.Through the monitoring video verification of the group pig house,it shows that the system can realize the functions of multi-target tracking of group pigs,recognition of feeding behavior of group pigs,and statistics of feeding behavior. |
PDF Full Text Request