Research On Photoelectric Detection And Tracking Algorithm For Pedestrian Targets Based On Security Unmanned Vehicles

With the combination and practical application of artificial intelligence technology in various industries,target detection and tracking for security scenarios have gradually become a research hotspot.In security scenarios such as warehouses and stations,intelligent patrol devices represented by security unmanned vehicles have become feasible solutions to improve the inadequate security patrol coverage and waste of security manpower resources.This study is based on the security unmanned vehicle scenario,fully utilizing pedestrian features and neural networks to achieve detection and tracking of pedestrian targets.Regarding pedestrian detection,due to the small number of pixels occupied by the target in infrared images and the weak features(small targets),this thesis takes the Yolov5 s 6.0version as the basic framework,adopts different attention mechanisms(SE,CBAM,CA),improved loss functions(CIo U),and optimized training data sets to enhance the network’s detection performance for weak infrared pedestrian targets.For pedestrian tracking,this thesis uses the Deep Sort network for multi-object tracking and uses the network weights trained in the detection stage for the pre-detection part of tracking,achieving stable tracking of pedestrian targets in real security environments.Experimental results data show that among different improvement methods for Yolov5 s network,the combination of improved loss function and added CA attention mechanism works best,and the detection accuracy of pedestrian target m AP0.5 improves by 0.2% to98.5% under LLVIP dataset compared to the control method.With the homemade fusion dataset,the detection accuracy of pedestrian target m AP0.5 improves by 0.4% to 97.6%compared with the control method,and the improvement is obvious for the detection of small targets in simulated security scenes compared with the model before improvement.The detection rate of weak targets is significantly improved without increasing the detection time,and the continuous tracking of weak targets can be achieved after loading the pre-trained weight model of Deep Sort network.In terms of detection speed,the improved target detection network takes an average of 10 ms per frame to detect the target,and the tracking network takes an average of 35 ms per frame,with a total duration of about 45 ms per frame,which can meet the real-time target tracking requirements of security scenes.It has some significance for the intelligent unmanned vehicle to land on the security scene.