A Lightweight Model for Detecting Dangerous Driving Behaviors in Road Transportation

In recent years, there has been a concerning surge in road traffic accidents, with hazardous driving behaviors among truck drivers identified as significant contributors. Consequently, accurately detecting and preventing these dangerous behaviors have become paramount. While convolutional neural networks (CNNs) are commonly used for driver behavior detection, challenges persist, particularly in detecting small-scale objects due to data quality and model architecture limitations. This research introduces an innovative approach utilizing the YOLOv8-SG model to detect dangerous driving behaviors. We have incorporated the SPD-G module to enhance the precision of detecting small-scale target behaviors. Additionally, we have developed the G2f module to optimize the model’s lightweight characteristics, meeting the performance demands of transport companies and their vehicular computing equipment. We also fine-tuned the loss function penalty terms and introduced the EXIoU loss function to improve boundary box regression performance. Empirical results show that our model outperforms the baseline, achieving a 1.1% mAP increase while reducing the model’s parameters by 14.3% and overall size by 12.9%. Notably, small-scale target detection, such as closed-eye behavior, improved by 3.8% in mAP, effectively reducing omissions and false positives.