Research on traffic supervision and special vehicle detection algorithm based on deep learning

Addressing vehicle detection requirements in complex traffic scenarios, this paper proposes an enhanced lightweight detection framework based on the CBAM attention mechanism for YOLOv8. By embedding the CBAM module within YOLOv8's backbone network and feature pyramid, a synergistic optimisation structure of "feature extraction-channel-space calibration-multi-scale fusion" is established, effectively enhancing the model's feature representation capabilities for small targets and occluded vehicles. A dedicated vehicle subset comprising buses, cars, and motorcycles was constructed using the PASCAL VOC 2012 dataset, encompassing vehicle samples across varying scales, angles, and lighting conditions. Experimental results demonstrate that the improved model achieves significant gains in detection accuracy, with an overall mAP50 of 83.3% and mAP50-95 of 63.5%, representing improvements of 3.3% and 1.8% respectively over the baseline model. Particularly in small object detection, the motorcycle category exhibits the most pronounced accuracy enhancement. Visualisation results further confirm that the improved model exhibits greater robustness and accuracy in complex backgrounds, occluded objects, and multi-scale scenarios, providing reliable technical support for intelligent transport systems and autonomous driving applications.