Road Traffic Target Detection Based on CA-YOLOv8

In recent years, automobile enterprises have continued to promote the automation and intelligence level of the whole vehicle, in which target detection is a crucial part of the auto drive system. However, the auto drive system has not yet achieved perfect processing in accurately sensing the road and locating surrounding targets, especially at night and in severe weather conditions. In order to overcome these obstacles, this work suggests an enhanced small object detection algorithm Coordinate Attention YOLOv8. First of all, as the auto drive system can not accurately sense the road and locate the surrounding targets, CA attention mechanism is adopted to integrate it with C2f module to form CA_ C2f, which is introduced into Backbone and Neck network to strengthen the target information and location information in the features and improve the model's capacity for feature fusion. Secondly, in severe weather conditions, detection accuracy may decrease. The enhanced model's loss function, which employs Wise IoU to adaptively modify weight coefficients, enhances both its durability of detection and bounding box regression performance. Finally, to address the loss of information caused by uneven scales in detecting small targets, to enhance the acquisition of the target's attributes and contextual data, a thin target detection layer is incorporated. According to the experimental results, the improved model's accuracy, recall rate, and average accuracy in the computer vision algorithm evaluation dataset KITTI in the automated driving scene are, respectively, 2.1, 10.8, and 6.9 percentage points higher than those of the original model. This effectively addresses the issue of the small target's poor detection effect in intricate traffic scenes, makes the auto drive system better improve its performance.