An improved YOLOv7 method for vehicle detection in traffic scenes

作者：Yanwei Wang, Ye Tian, Junting Cheng, Xianglin Meng, Zeming Xie · 年份：2023 · DOI：10.1109/ccdc58219.2023.10326522 · 被引用次数：6 · 研究领域：Advanced Neural Network Applications、Advanced Data and IoT Technologies、Video Surveillance and Tracking Methods

Vehicle target detection is a key task in intelligent traffic management and has important significance for the construction of intelligent transportation systems. To address the problem of the adaptability of target detection algorithms in actual traffic scenes, a vehicle target detection algorithm based on the improvement of YOLOv7 is proposed. Firstly, the original backbone network is replaced with the MobileNetV3 network with a depth separable convolution structure, which reduces the model parameters and computational complexity. Secondly, the BiFPN structure is introduced in the feature fusion module to combine the position information of low-level feature maps with the semantic information of high-level feature maps. Finally, the Focal-Loss EIOU is used as the positioning loss function to accelerate the convergence speed of the predicted box. Based on the UA-DETRAC dataset, the best improvement strategy is determined through comparative experiments, and the superiority of the combination of improvement strategies is verified through ablation experiments. The experimental results show that the improved YOLOv7 algorithm has an average precision of 93.1%, which is 6.6% higher than the original YOLOv7, and the frame rate reaches 41, meeting the real-time requirements.