Rapid, accurate and robust detection of road vehicles is a core and fundamental task in fields such as intelligent transportation systems, autonomous driving and traffic management. In response to this demand, this paper proposes an improved YOLOv8 model, which optimizes feature extraction and scale adaptability by introducing the MBConv module and the C2fAKConv structure. The improved model has achieved a significant improvement in overall detection performance compared to the original model: the average accuracy (mAP50-95) has been greatly enhanced from 0. 352 to 0. 466. The recall rate (R) increased from 0. 421 to 0. 539, significantly reducing the rate of missed detections. The precision rate (P) has increased from 0. 632 to 0. 735, effectively reducing false detections. In terms of specific categories, the improvements are particularly prominent: the recall rate of the Car category has jumped from 0. 252 to 0. 328, and the mAP50 has increased from 0. 392 to 0. 474, enhancing the ability to recognize occlusions and small targets. The mAP50-95 of the Truck category has been upgraded from 0. 183 to 0. 264, improving the detection of vehicles with complex shapes. The Ambulance category maintained a mAP50 above 0. 9 while MAP50-95 was raised to 0. 799, achieving more accurate bounding box localization. The comprehensive results show that the proposed improvement strategy effectively enhances the feature extraction ability, scale adaptability and detection robustness for weak texture targets of the model. This research provides a more effective solution for real-time target detection in complex traffic scenarios.
Ting Zhang (Tue,) studied this question.