Vehicle Recognition and Driving Information Detection with UAV Video Based on Improved YOLOv5-DeepSORT Algorithm

• Published in: Sensors (Basel, Switzerland) Vol. 25; no. 9; p. 2788
• Main Authors: Zheng, Binshuang, Zhou, Jing, Hong, Zhengqiang, Tang, Junyao, Huang, Xiaoming
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI AG 28.04.2025; MDPI
• Subjects: Accuracy; Algorithms; Artificial intelligence; Behavior; Cameras; Deep learning; Drone aircraft; Methods; Monitoring systems; Neural networks; ramp; Remote sensing; Researchers; Sensors; Software; Traffic; UAV video; vehicle recognition; vehicle track information; Vehicles; YOLOv5 algorithm; China; YOLOv5 algorithm; vehicle track information; UAV video; vehicle recognition; ramp
• ISSN: 1424-8220; 1424-8220
• DOI: 10.3390/s25092788

To investigate whether the skid resistance of the ramp meets the requirements of vehicle driving safety and stability, the simulation using the ideal driver model is inaccurate. Therefore, considering the driver’s driving habits, this paper proposes the use of Unmanned aerial vehicles (UAVs) for the collection and extraction of vehicle driving information. To process the collected UAV video, the Google Collaboration platform is used to modify and compile the “You Only Look Once” version 5 (YOLOv5) algorithm with Python 3.7.12, and YOLOv5 is retrained with the captured video. The results show that the precision rate P and recall rate R have satisfactory results with an F1 value of 0.86, reflecting a good P-R relationship. The loss function also stabilized at a very low level after 70 training epochs. Then, the trained YOLOv5 is used to replace the Faster R-CNN detector in the DeepSORT algorithm to improve the detection accuracy and speed and extract the vehicle driving information from the perspective of UAV. By coding, the coordinate information of the vehicle trajectory is extracted, the trajectory is smoothed, and the frame difference method is used to calculate the real-time speed information, which is convenient for the establishment of a real driver model.