Pig Multi-target Recognition And Tracking Based On YOLOv4+DeepSORT

With the continuous improvement of my country’s economic efficiency,people’s requirements for food quality are increasing.Pork is an indispensable food,and it is imperative to promote the level of large-scale pig breeding.This paper uses the multi-target group pig raising video in the pig house as the data source.First,the identity of the pigs raised in the group is detected.On this basis,the target tracking algorithm is used to track the pigs in the group to obtain a trajectory map,which is combined with the diet of the pigs The corresponding behavior duration and frequency of the five behaviors of,drinking,rest,excretion and activity are abnormal judgments for the health status of the multi-objective pigs in the pig house.The main research work of this paper is as follows:(1)Taking the group raising pigs in the pig house as the data source,in-depth analysis of the methods used by domestic and foreign researchers to identify and detect pigs,summarize the application of machine vision-based livestock and poultry tracking,and analyze abnormal behaviors of pigs.The research status is analyzed.(2)In terms of pig identification and detection,the theory of Centernet,SSD,Faster R-CNN and YOLOv4 model methods are explained,the four models are compared and analyzed,and the most suitable algorithm for pig target detection is selected based on experimental analysis.YOLOv4 was used to identify multiple targets for group-raising pigs.The experimental results showed that the AP(average accuracy of each category)value of the YOLOv4 model on the test set for the identification of 8 pigs was 99.94%,98.45%,and 99.70%.,98.84%,99.46%,99.15%,99.99%,99.45%,and the overall m AP(average accuracy rate of all categories)reached 99.4%,which is an increase of 2.7 percentage points,3.3 percentage points,and 7.0 respectively over CenterNet,SSD,and Faster R-CNN Percentage points.The detection speeds of CenterNet,SSD,Faster R-CNN and YOLOv4 on the test set are 27 FPS,15FPS,3FPS and 48 FPS respectively.The experimental results show that YOLOv4 can accurately and quickly perform target detection and identity recognition on group pigs.According to the experimental data It can be seen that YOLOv4 is the best in real-time detection.Only by completing the identification and detection of pigs can the tracking trajectory of the pig target be finally obtained.(3)The traditional multi-target pig tracking algorithm in target tracking has low accuracy and is easy to lose pig targets.In order to further improve the accuracy of tracking,this paper adopts a combined model that combines YOLOv4 and DeepSORT.The model first uses YOLOv4 to identify and detect group pigs,and then uses YOLOv4 to correlate with DeepSORT to track group pigs and draw a motion trajectory to improve the accuracy of pig target tracking.Using this model to verify on the VOC data set,the MOTA index of this model is higher than the CenterNet+DeepSORT model,SSD+DeepSORT model and Faster R-CNN+DeepSORT model.Experimental results show that the model can accurately track multi-target pigs and provide a basis for judging the health of pigs.(4)According to the track map obtained by tracking combined with the analysis of pigs in the pig house environment,the areas are divided according to the pig farm environment,and five areas are divided into eating area,drinking area,rest area,excretion area,and activity area.According to the motion trajectory diagram,a statistical table is established for the duration and frequency of each pig’s behavior,and the abnormality of the pig is judged according to the data in the statistical table and other documents.