Research On Multi-objective Identification And Tracking Technology For Live Pigs Based On Improved Deep SORT

With the development of the modern breeding industry,pig production management is becoming more and more welfare-oriented,systematic and intelligent.Accurate identification and tracking of individual pigs,as well as behavioral discrimination and statistics of pigs,are important components.Traditional pig identification and tracking methods are time-consuming and laborious,which is not conducive to the refined breeding of pigs.In recent years,with the development of deep learning,the modern breeding industry has gradually adopted neural networks for the non-invasive identification of individual pigs.In this thesis,deep learning algorithms are used to realize multi-target detection,tracking,behavior discrimination and live pig statistics.The details are as follows:(1)This thesis presents an improved YOLOv5 model that introduces a pig head and neck recognition method.To enhance the model’s accuracy,we build on the YOLOv5 algorithm and implement the following improvements: First,we replace the Euclidean distance of K-Means clustering with 1-IOU to improve the fitness of the model’s target box.Next,we incorporate a coordinate attention mechanism in the backbone network to learn small target features and locations more effectively.Lastly,we introduce Bi FPN feature fusion in the neck to expand the model’s receptive field and enhance multi-scale learning for disturbance targets.These changes result in a 2.2% increase in accuracy for pig head and neck recognition under normal circumstances.Moreover,our model also improves the missed detection situation in pig-intensive scenarios and the missed detection and low classification confidence in long-distance small target scenarios.(2)This thesis introduces the improved Deep SORT model’s pig multi-target tracking algorithm.The Deep SORT algorithm,which has better tracking capabilities,is used as a benchmark,and three improvements are made to it.Firstly,the detector is replaced with the optimal YOLOv5 improved model11,which improves the detection accuracy.Secondly,the re-identification network model is replaced to enhance the algorithm’s appearance information acquisition ability.Finally,the trajectory generation and matching process of the Deep SORT algorithm are optimized,and a second IOU matching is added to reduce ID mismatch and tracking mismatch.After screening the experimental video sequence,we selected five videos with varying degrees of complexity for testing.The improved Deep SORT algorithm demonstrated varying degrees of improvement when compared to the original tracking algorithm,as measured by IDS,IDF1,MOTA,and MOTP.Notably,IDS was reduced by 63.1%,and MOTA was increased by 1.3%.These improvements are particularly beneficial for tracking the head and neck of pigs,which lays a stronger foundation for their overall health.(3)This subject thesis discusses the introduction of a pig health care system that utilizes multi-target detection and tracking of pigs,as well as pig behavior discrimination and statistics.To achieve pig behavior discrimination,two-dimensional coordinates are used to differentiate between drinking water and diet,while three-dimensional coordinates are used to distinguish between high-active,general-active,and inactive behavior.The system also counts the number of pig behaviors and the total duration.The system is supported by front-end and back-end codes and hardware,enabling multiple modules such as pig head and neck detection,pig multi-target tracking,and pig behavior recognition and statistics.The result is a comprehensive pig healthcare system that can serve as a reference for practitioners in the field.