Lightweight Traffic Object Detection Model Based on Improved YOLOv5n

Engineering notes

The results show that the improved model achieves 86.24% mAP@0.5, 98 FPS inference speed, and only 2.23 M parameters. Compared with the original model, the accuracy is significantly improved.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Aiming at the problems of the lightweight YOLOv5n model in complex traffic scenarios, such as easy missed detection of small objects, low recognition accuracy of occluded objects, and loss of feature fusion information, this paper proposes an improved detection model with multi-module collaborative optimization. Firstly, the original CSPDarknet53 backbone network is replaced and fine-tuned with the HGNetV2 lightweight structure to enhance the feature extraction ability for small-scale and occluded objects. Secondly, the Detect_MBConv module is introduced to reconstruct the detection head, reducing computational overhead based on depthwise separable convolution. Meanwhile, Dysample dynamic upsampling is used to replace traditional interpolation to improve the quality of cross-scale feature fusion. Experiments are conducted based on the BDD100K traffic dataset. The results show that the improved model achieves 86.24% mAP@0.5, 98 FPS inference speed, and only 2.23 M parameters. Compared with the original model, the accuracy is significantly improved. The model can be effectively applied to edge low-computing devices such as vehicle-mounted terminals and surveillance cameras, meeting the requirements of real-time and high-precision detection in actual traffic scenarios.

Links and sources

• PDF from original source

Comments