Research On Vehicle Detection And Tracking Algorithms Based On UAV Images

Vehicle detection and tracking are critical components of intelligent transportation systems.Existing experiments mainly use fixed camera devices to capture vehicle video information,but this method has disadvantages such as poor flexibility and limited field of view.Unmanned aerial vehicles(UAVs)are widely used in various fields due to their efficient and flexible data collection methods.Applying UAVs to the field of intelligent transportation can quickly and accurately obtain vehicle information and improve work efficiency.However,due to the small target size and complex background in UAV images,the performance of existing object detection and tracking algorithms is not ideal.Therefore,this paper focuses on how to improve the performance of object detection and tracking algorithms,and the main research contents are as follows:(1)To improve image quality,image processing techniques are used to address image blur and lighting problems caused by UAV jitter and shooting environments in the Vis Drone2019 dataset.Bilateral filtering and contrast-limited adaptive histogram equalization techniques are used to smooth the image,remove noise,and improve contrast to improve image clarity and visibility,thereby reducing the difficulty of subsequent object detection and tracking tasks.(2)This paper proposes an improved YOLOX-s detection algorithm to address the low detection accuracy and large model parameter issues in vehicle detection using the YOLOX-s algorithm.The Sim AM attention module is added to the backbone network to improve the feature extraction ability of the backbone network,making it more accurate in detecting small targets in UAV images.The Bi FPN structure is introduced in the feature fusion stage to fully fuse features of different levels to obtain more comprehensive and accurate feature representations.The lightweight Ghost module is used to replace the ordinary convolution module of the backbone network to reduce model parameters and improve algorithm speed.The Focal Loss loss function is used instead of the cross-entropy loss function used for confidence loss to solve the problem of imbalanced positive and negative samples in the dataset,thereby improving the accuracy of object detection.Experimental results show that the improved YOLOX-s algorithm has an average precision improvement of 9.17% compared to the YOLOX-s algorithm,and the model parameter quantity is reduced by 1.7M.(3)This paper proposes an improved DeepSORT tracking algorithm to address the low tracking accuracy and frequent ID changes due to object occlusion in vehicle tracking using the Deep SORT algorithm.The improved YOLOX-s detector is used instead of the original detector in Deep SORT to improve the accuracy of vehicle detection.The extended Kalman filter is used instead of the Kalman filter to enhance the ability of Deep SORT algorithm to predict vehicle states in nonlinear environments.The CIOU distance metric is used instead of the IOU as the distance measure to improve the matching degree between detection boxes and tracking boxes and solve the problem of frequent ID changes due to long-term object occlusion.Experimental results on different video sequences show that the improved Deep SORT algorithm has improved MOTA and MOTP indicators compared to the original Deep SORT algorithm,and the number of object ID changes has correspondingly decreased,verifying the effectiveness of the proposed algorithm in this paper.