YOLO-SGC: A Dangerous Driving Behavior Detection Method With Multiscale Spatial-Channel Feature Aggregation

Abstract

In intelligent transportation system, it is significant to detect drivers’ dangerous driving behaviors accurately and in real time. However, current fatigue driving detection methods only focus on facial expressions or hand movements and ignore the impact of global behavior, resulting in poor detection results in complex scenes. To solve this problem, we propose a novel dangerous driving detection method called YOLO-SGC. This method builds on the YOLOv8 framework, enhancing it with the addition of spatial and channel reconstruction convolution (SCConv) and a global attention mechanism (GAM) to efficiently capture multiscale features based on spatial-channel information. Additionally, we leverage the cross-scale partial connections (CSPCs) method to optimize the spatial pyramid pooling fast (SPPF), expanding the model’s field of view. These innovations significantly enhance the algorithm’s ability to express features at various scales, thereby improving the model’s capacity to recognize multiscale driving behaviors. In addition, we construct a new dataset of real driving scenarios called QIN_Dataset, which contains 34 000 images of real driving scenes from frontal and side angles. Finally, we validated YOLO-SGC and other mainstream methods on the public datasets and our QIN_Dataset. The results show that compared with other mainstream models, YOLO-SGC has a significant improvement in the average accuracy (overall increasing 1%~42.7% in mAP50 and 1.1%~55.6% in mAP50-95) while maintaining a high detection speed. This demonstrate that YOLO-SGC is an effective and practical solution for vision-based driver dangerous behaviors detection.

Topics & Concepts