Abstract: In view of the problems of low detection accuracy and high missed detection rate in vehicle and pedestrian detection at roadside ends, a vehicle and pedestrian detection algorithm YOLOv8-APC based on attention scale sequence fusion was proposed. Firstly, the scale sequence fusion module (SSFF) and the three-feature encoder (TFE) were used in the neck network to enhance the extraction and fusion of multi-scale information, meanwhile, the channel and position attention mechanism (CPAM) was introduced to improve the detection accuracy. Then, the P2 detection layer was added on the basis of the improved network structure to improve the detection ability of small targets and reduce the missed detection rate. Finally, the C2f_GhostDynamicConv (C2f_GDC) module was applied in the backbone network to effectively reduce the complexity of the model. To verify the effectiveness of the proposed algorithm, the validation was conducted on the roadside end dataset Vapddsits in the Chongqing Science Valley Demonstration Zone. The experimental results show that the mAP50 value and recall rate of YOLOv8-APC are 11.1% and 11.9% higher than those of the original model; the parameter quantity and model volume are only 1.85M and 4.1MB respectively, which are 38.3% and 34.9% lower than those of the original model. The proposed algorithm can achieve more accurate detection of distant small targets and occluded targets, which doesn’t occupy too much memory resources, providing a solution for vehicle and pedestrian detection at roadside ends.

Key words: traffia and transportation engineering; YOLOv8; vehicles and pedestrians; feature extraction; attention mechanism; scale sequence fusion

摘要： 针对在路侧端车辆与行人检测中存在检测精度低，漏检率较高等问题，提出了一种注意力尺度序列融合的车辆行人检测算法YOLOv8-APC。首先，在颈部网络中使用尺度序列融合模块SSFF与三特征编码器TFE，以增强对多尺度信息的提取与融合，同时引入通道与位置注意力机制CPAM提高检测精度。然后，在改进后的网络结构基础上增加P2检测层，提高对小目标的检测能力，降低漏检率。最后，在主干网络中应用C2f_GhostDynamicConv (C2f_GDC)模块，有效降低模型的复杂度。为验证算法的有效性，在重庆科学谷示范区路侧端数据集Vapddsits上进行验证，实验结果表明：YOLOv8-APC的mAP50值与召回率较原模型提升了11.1%、11.9%；参数量与模型体积分别仅有1.85 M、4.1 MB，分别较原模型下降了38.3%、34.9%，其对远距离小目标以及遮挡目标能够实现更为准确的检测，且不会占用过多的内存资源，为路侧端车辆行人检测提供了一种解决方案。

关键词: 交通运输工程；YOLOv8；车辆与行人；特征提取；注意力机制；尺度序列融合