Research On Real-time Vehicle Detection And Tracking Algorithm Based On UAV

Due to its advantages of flexibility,high efficiency and low cost,UAV is increasingly popular in navigation search and rescue,traffic monitoring and other aspects.UAV video monitoring played an important role in the field of security and prevention.The realization of target detection and tracking from the perspective of UAV can effectively improve the efficiency and intelligent operation level of UAV.This article studied the high rate of vehicle missing detection and the difficulty to meet the real-time requirements of the embedded platform from the perspective of UAV,and designed a real-time detection and tracking algorithm of multi-target vehicle applied to the UAV platform to solve the problems of missing detection,false detection and occlusion in the detection and tracking from the perspective of UAV.The main work was as follows:(1)Aiming at the problems of low accuracy and poor real-time performance of the existing target detection network on the UAV platform,four improvement strategied were proposed based on YOLOv3.Firstly,the basic feature extraction layer was constructed by using the anti residual network,and the Mobilenetv2 was used as the backbone network for structural lightweight processing;secondly,in order to solve the problem of inaccurate vehicle positioning and unbalanced sample,the DIoU and Focal Loss were used instead of position loss and confidence loss in the original loss function,and obtained the anchor size through k-means clustering algorithm.Finally,the data set was enhanced by the fast ACE algorithm to improve the performance.The experimental results showed that mAP of the improved network was improved by 4.54%,and FPS was increased from 6 frames/s to 14 frames/s.(2)On the basis of detection,aiming at the shortcomings of sort algorithm,such as ID switch,etc.An improved Kalman filter multitarget tracking algorithm was proposed,which took the real-time HSV feature and HOG feature as the appearance information of the target,used the linear weighting of the motion information and appearance information based on Mahalanobis distance as the fusion correlation measure,and used the Hungarian algorithm to predict and match the target trajectory.The experimental results showed that the improved algorithm had a good track assignment effect for long-term occlusion,target re-entry and so on.The MOTA was increased by 4.5%,and the overall tracking effect was significantly improved.(3)TensorRT was used to design the reasoning acceleration optimization framework for the detection model.We deployed the optimized detection and tracking algorithm in NVIDIA Jetson TX2 system,and carryed out the tracking experiment for the actual scene taken by the UAV.The test results showed that the tracking accuracy loss was very small and the speed was increased to 22 frames/s.The algorithm proposed in this article could accurately and continuously identify and track multiple vehicle targets,met the requirements of real-time and accuracy for UAV platform detection and tracking,and provided a set of prototype system for later application in intelligent security and intelligent UAV products.