关键词: YOLOv12；注意力机制；多尺度特征提取；目标检测

Abstract: Aiming at the problems of missed detection and false detection caused by diverse target scale variations, complex background interference and mutual overlap between targets in complex traffic scenarios, this paper proposes YOLO-PCA, an improved road object detection algorithm based on YOLOv12. First, the poly kernel inception(PKI) module is introduced into the backbone network to construct a multi-scale feature extraction module. Through the combination of convolution kernels of different sizes, features of different scales are extracted, which enhances the feature representation ability of the model for targets of different scales. Second, the channel prior convolutional attention (CPCA) module is introduced to dynamically allocate attention weights in channel and spatial dimensions, which enhances the model's ability to extract detailed target feature under complex backgrounds while reducing computational complexity. Finally, the parallelized patch-aware attention (PPA) module is introduced into the detection head, enabling the model to better focus on key information of small targets and improves the detection accuracy of the model for small targets. Experimental results show that the mean average precision (mAP@0.5) of the improved model on the KITTI dataset reaches 98.7%, which is 0.9 percentage points higher than the baseline YOLOv12 model. Therefore, the improved model effectively improves the detection performance for occluded, truncated targets and small targets, and can better meet the requirements of complex road object detection.

Key words: YOLOv12; attention mechanism; multi-scale feature extraction; object detection