Research on visual vehicle detection and tracking based on deep learning. Issue 1 (July 2020)

• Record Type: Journal Article
• Authors: Zhang, Yaoming; Song, Xiaoli; Wang, Mengen; Guan, Tian; Liu, Jiawei; Wang, Zhaojian; Zhen, Yajing; Zhang, Dongsheng; Gu, Xiaoyi
• Is Part Of: IOP conference series. Volume 892:Issue 1(2020)
• Publication Date: 2020-07
• Subjects: Materials science -- Periodicals; 620.1105
• DOI: 10.1088/1757-899X/892/1/012051
• Languages: English
• ISSNs: 1757-8981

Abstract

Abstract: At present, vehicle tracking technology is still a hot research direction all over the world. The main work of this paper is divided into two parts: we should complete visual vehicle detection based on deep learning and vehicle tracking based on similarity measurement and association algorithm. The main work of this paper is as follows: Based on the YOLOV3 detection algorithm, this paper puts forward some new ideas and ideas, which are applied to the video detection and experimental results. Based on the structure of YOLOV3 convolutional neural network, the vehicle video detection experiment is carried out under the framework of TensorFlow, and the detection results are evaluated by the detection accuracy and error detection rate. Firstly, the augmented data set of video detection is made, and a section of driving record video is recorded under the surrounding road conditions of the school. After that, the data set is used to train the weightless network of YOLOV3. Finally, the test results between KITTI data set and augmented data set are compared. Then the network is used to compare the detection results of different detection algorithms. The second work is to build the vehicle tracking model after getting the object selection frame after video detection. In this model, the classic hog feature extraction method is used to extract the vehicle appearance features, and then the motion similarity is calculated. After getting the total similarity between the targets in the front and back images, the target trajectory is recovered or released by the target trajectory management algorithm to achieve the tracking effect. The above vehicle tracking model is optimized by three evaluation parameters: MOTA, MOTP and IDS. In the end of this paper, the tracking effect experiments under different detection algorithms, different detection thresholds and the peeling analysis of tracking features are carried out for the vehicle tracking part, and some conclusions are obtained.