Abstract: The detection of road traffic participants such as pedestrians, cyclists and vehicles is regarded as one of the core tasks for achieving autonomous driving. However, in complex scenarios such as uneven lighting, occlusion, dense targets, and small distant objects, false detections and missed detections frequently occur. To address these issues, an improved YOLOv8 object detection algorithm for complex traffic scenes was proposed. Based on the lightweight GhostNet network structure, the backbone and neck networks of the YOLOv8 model were optimized. Standard convolutions (Conv) were replaced by ghost convolutions (GhostConv), while the C2f module was substituted with a combination of the ghost bottleneck (G-bneck) and the C3 module, effectively suppressing redundant detections and improving detection efficiency. A mixed local channel attention (MLCA) mechanism was employed to integrate multi-scale information, enhancing the feature extraction capability of the model. Additionally, a small-object detection layer was incorporated to preserve more detailed features, thereby improving detection performance for small distant targets. Finally, the wise intersection over union (WIoU) loss function was adopted to accelerate network convergence and enhance robustness in complex scenarios. Experimental results demonstrate that the improved YOLOv8 model achieves a mean average precision of 0.872 on the constructed RCCW-Dataset for complex traffic scenes, representing a 2.1% improvement over the original model. Meanwhile, the number and size of model parameters have been reduced by 41% and 37% respectively, enabling effective detection of target tasks in real-time complex traffic scenarios.

Key words: traffic engineering; YOLOv8 model; lightweight; MLCA attention mechanism; WIoU

摘要： 对道路交通参与者中的行人、骑行者以及车辆进行检测是实现自动驾驶的核心任务之一。在光照不均、遮挡、密集目标和远距离小目标等复杂场景中往往会存在误检及漏检情况；基于此，提出了一种改进YOLOv8模型的复杂交通场景目标检测算法。基于GhostNet轻量化网络结构，对原始YOLOv8模型的主干和颈部网络进行改进，利用幻影卷积(ghost convolution, GhostConv)来替换标准卷积(convolution, Conv)，并将幻影瓶颈(ghost bottleneck, G-bneck)结合C3模块代替C2f模块，这样就有效抑制了冗余检测，提升了检测效率；应用混合局部通道注意力机制(mixed local channel attention, MLCA)对多元化信息进行整合，以增强模型的特征提取能力；添加小目标检测层，可保留更多细节特征信息，提高对远距离小目标的检测能力；采用WIoU(wise intersection over union loss)损失函数加速了网络收敛并增强了在复杂工况下的鲁棒性。研究结果表明：改进YOLOv8模型在所构建的复杂交通场景数据集RCCW-Dataset中的平均精度均值为0.872，较原始模型提高了2.1%；模型参数量和大小分别降低了41%和37%，能有效完成对实时复杂交通场景目标任务的检测。

关键词: 交通工程； YOLOv8模型； 轻量化； MLCA注意力机制； WIoU