Application Research On Improved Hungarian Algorithm For Trajectory Tracking And Intelligent Behavioral Identification Of Stored Grain Pests

Grain pests are pests that live in grain depots,grain processing plants and other grain storage sites.They harm grain,oilseeds and processed products.A large number of stored grain pests adversely affect the quality and quantity of food,long-term use of chemical agents to kill insects brings environmental pollution and other problems.And people usually kill insects by using heat treatment in the processing plant to kill insects,if the temperature is not properly controlled,the pests will be killed by the high temperature before it leaves the machine,and the insect debris will damage the machinery.Therefore,it is necessary to find a safe and efficient storage pest control scheme instead of chemical control.This paper explores the biological characteristics of stored grain pests by studying the movement behavior of stored grain pests at different temperatures.Specifically,this paper tracks the trajectory of stored grain pests using an artificial intelligence method,and for the problem of object overlap in multi-object tracking,a Hungarian algorithm incorporating cosine similarity combined with the LSTM network prediction method is proposed to achieve effective tracking of overlapping objects,and the effectiveness of the method is verified by experiments.The main work of this paper is as follows:(1)Improve the accuracy of object detection.On the basis of YOLOv5 x network,the SE attention module is added to learn the global channel characteristics,and then the number of detection layers is appropriately reduced according to the size of stored grain pests in the picture,so as to reduce model parameters and improve training speed,and finally introduce the SIo U loss function to accelerate the matching speed between the prediction box and the real box,so that the model converges faster.(2)Improve tracking accuracy for overlapping multi-objects.Aiming at the problem that it is difficult to effectively track when the objects overlap,this paper draws on the design ideas of adding a re-identification network to the Deep Sort algorithm,uses the LSTM network to learn the trahjectory characteristics of overlapping objects,assigns their respective IDs after the overlap,reduces the frequency of ID switch,and integrates cosine similarity into the Hungarian algorithm to associate with the original trajectory,so that it can continue to track accurately.Finally,the trajectory tracking and behavior statistics results are visualized in real time.In this paper,260 motion videos of stored grain pests were collected,including75 single-target videos,87 multi-target videos without overlap,and 118 overlapping multi-target motion videos.The detection accuracy of stored grain pests using the improved network named YOLOv5x＿improved reached 98.4%,which was 3.2percentage points higher than that of the original YOLOv5 x network.The improved Hungarian algorithm combined with LSTM network achieves more than 98% video tracking accuracy in videos without overlap,while the average multi-object tracking accuracy m MOTA and video tracking accuracy of the method reach 0.9904 and 92.4%,respectively,which are higher than Hungarian algorithm and Deep Sort and other algorithms.The improved Hungarian algorithm in this paper,specifically the Hungarian algorithm integrating cosine similarity combined with LSTM network prediction method,has a good tracking effect in the trajectory tracking task of grain storage pests,which establishes a higher credibility for the study of the biological behavior characteristics of stored grain pests and provides a reference for the establishment of efficient storage pest control strategies.