Vehicle Analysis System Based on DeepSORT and YOLOv5
Fuheng Guo, Yi Xu · 2022 3rd International Conference on Computer Vision, Image and Deep Learning & International Conference on Computer Engineering and Applications (CVIDL & ICCEA) · 2022
Nowadays, the constantly expanding coverage of expressways leads to an astronomical volume of daily high-speed traffic. In order to address the poor real-time performance of traditional traffic statistics and low accuracy in detection of densely distributed small targets for original algorithms, this paper proposes a vehicle analysis system that is based on DeepSORT and YOLOv5. More specifically, YOLOv5 is adopted to detect the target, Kalman filter is employed for prediction and update, and the Hungarian algorithm is applied for trajectory matching in the cascade matching, so as to realize the multi-target tracking of vehicles. Besides, inverse perspective mapping is performed to obtain the actual distance traveled by a vehicle per unit time, and to calculate the speed of it. The experimental results show that the vehicle analysis system based on DeepSORT and YOLOv5 has an average accuracy of 96% for vehicle targets, a 98.36% accuracy of vehicle speed measurement, and a lower model loss of than 2% on average, which meets the needs of real-time analysis.