Driving risks from light pollution: an improved YOLOv8 detection network for high beam vehicle image recognition

The effective detection of high beam vehicles is useful to reduce the occurrence of traffic accidents and light pollution at night. It is time-consuming and laborious to detect the high beam vehicles by human. We propose a model Global Attention Content-Aware ReAssembly of FEatures Small Target Detection based on YOLOv8 (GCS-YOLOv8) to improve the efficiency of high beam vehicles detection. Firstly, a cross-stage feature fusion global attention module (CGM) is designed in the backbone to reduce the information loss during down-sampling and magnify global dimension-interactive features. Secondly, a Convolution-1*1 Content-Aware ReAssembly of FEatures Module (CCM) is designed in the neck to reconstruct subtle feature maps of high beam vehicles. Meanwhile, a CCM Small Target Detection Module (CSM) is designed in the neck to derive the feature maps with small receptive field. Finally, we set up a high beam vehicles dataset. The model GCS-YOLOv8 achieves the mean average precision (mAp) of 94% on the test set, and it is higher than other models. In addition, multiple ablation experiments are done to prove the effectiveness of our model GCS-YOLOv8.