Research on a Lightweight and Efficient Vehicle Detection Approach Based on YOLO Algorithm

Abstract

In recent years, intelligent transportation systems have developed rapidly, and autonomous driving systems, as key solutions to traditional traffic problems, have received increasing attention. Vehicle and pedestrian detection serves as a fundamental component within these systems. However, in practical application scenarios, it is challenging to achieve a good balance between algorithm accuracy and resource consumption; furthermore, accuracy degrades under complex conditions such as occlusion and illumination variations. To address these challenges, this paper proposes an efficient and lightweight vehicle and pedestrian detection model named YOLO-MDT, based on the YOLOv8n architecture. Firstly, we introduce a Feature Dynamic Task-Sharing Detection Head (DT-Head). This head reduces the number of parameters through weight sharing and enhances detection performance by facilitating feature interaction between classification and localization tasks. Secondly, we design a C2f-MEE module to replace the original C2f module. This new module improves multi-scale feature extraction and edge awareness capabilities, thereby boosting object detection performance. Finally, DySample upsampling and the Focal-PIoUv2 loss function are employed to replace the original upsampling method and loss function, respectively. This improves model convergence speed, enhances bounding box accuracy, and reduces computational overhead. The proposed model was validated on the SODA10M and KITTI datasets. It achieves a parameter count that is only 70% of the original network and reduces computational complexity by 0.1 GFLOPs. Significant improvements in mean Average Precision (mAP) were observed: mAP50 increased by 4.3% and 3.1%, while mAP50-95 increased by 2.9% and 3.2% on the SODA10M and KITTI datasets, respectively.