Heterogeneous Multi-agent Target Collaborative Recognition And Formation Tracking

In tasks such as epidemic prevention and rescue,drones and unmanned vehicles have been widely used,and they play an irreplaceable role in the efficiency and reliability of completing tasks.In some complex environments,a single drone or unmanned vehicle has some limitations in terms of execution mode,action space,etc.when completing tasks,which will seriously affect its ability to complete tasks.Secondly,in some target detection scenarios,if the number of targets is too dense or overlaps,it can cause the target detection algorithm to miss and misdetect,affecting the execution of subsequent tasks.In the process of target tracking,when the target moves too fast or is obstructed,the target tracking algorithm loses the target and the tracking fails.In response to the above issues,the main research content of this article is as follows:(1)Propose a heterogeneous intelligent agent collaborative recognition and formation tracking system.This article designs a collaborative system that includes three parts:drones,unmanned vehicles,and ground control consoles.There is a strong complementarity between drones and unmanned vehicles.Drones have a wide field of vision and fast movement speed,which can track targets well and send target information and position to the unmanned vehicle,expanding its perception range.Unmanned vehicles have high definition and reliable operation.(2)The attention mechanism and loss function are introduced to improve the target detection algorithm and improve the detection accuracy and robustness in UAV.The added attention mechanism is a SE(Squeeze and Extraction Networks)module based on a weight matrix,which enhances the detection ability of YOLOv5s(You Only Look Once version 5 small)by extracting important information from the image and reducing missed and false detections.This module can extract different weight sizes based on different positions,thereby improving the model’s adaptability to image positions.The replaced loss function includes taking the length and width of the detection frame as one of the important factors for detection according to the shape characteristics of the object,and integrating many other factors,including the intersection of the prediction frame and the real target frame and the distance of the center point.The loss function is improved so that the gradient can change adaptively with the distance between two frames,which is more conducive to the network training process.(3)Solve the problem of target occlusion loss during target tracking and improve the robustness of unmanned vehicle tracking.Integrating target detection algorithm and target tracking algorithm,the tracking algorithm uses the detected results as the tracking target.Outlier detection is introduced to determine whether the target is lost because the target is blocked or not in the field of vision.On the basis of the KCF-DSST(Kernel Correlation Filter Discriminative Scale Space Tracker)target tracking algorithm,a outlier detection module is added to determine whether the target tracking is abnormal.At the same time,a target re detection mechanism is introduced to identify lost targets and reposition them.