关键词: 关键词: 改进YOLOv8, 不安全驾驶行为检测, 深度学习, 目标检测算法

Abstract: In order to effectively detect unsafe behaviors, such as drivers not wearing seatbelts and looking at cell phones in real time under various complex lighting conditions, a deep learning-based method for detecting unsafe driving behaviors under complex lighting has been designed. The YOLOv8n model has been selected as the basis, and a series of targeted improvement measures have been proposed to enhance its detection performance. First, the P6 scale is increased to allow the model to capture the diversity of unsafe driving behaviors under various lighting more comprehensively.Then, a spatially separable adaptive convolutional module (SSAC) is used to replace the traditional convolutional module of the backbone network, improving the feature extraction accuracy while achieving lightweighting.Next, Channel Prior Convolutional Attention (CPCA) is introduced to effectively enhance the network's attention to important features and improve feature representation. Finally, the original YOLOv8n neck network is replaced using the Selective Attention Feature Fusion (SAFF) structure to further enhance the comprehensive performance of the model. The experimental results demonstrate that the enhanced YOLOv8n model enhances the mean average precision (mAP) from 0.969 to 0.990 compared to the original version, representing a 2.17% improvement. Under normal lighting conditions, the model exhibits an enhancement of 1.76%. In night scenes, the improvement is 1.75%. In the backlight environment, the enhancement is 2.42%. In addition, the enhanced YOLOv8n model demonstrates a substantial advantage over other models (e.g., YOLO11n, RT-DETR) with an FPS of 118, exhibiting a synergy between accuracy and processing speed.

Key words: Keywords: improved YOLOv8, unsafe driving behavior detection, deep learning, target detection algorithm