Abstract: To improve the detection accuracy and recognition precision of traffic lights, this paper proposes an SVDD-YOLOv8 target detection method based on YOLOv8. This method enhances feature capture and reconfirms targets by integrating the global attention mechanism (GAM) and support vector data description (SVDD) classification module. Meanwhile, the embedding intersection over union (EIoU) loss function is introduced to improve positioning accuracy. The abnormal areas identified by SVDD are retrained to enhance the model's performance. Experiments show that this method improves detection accuracy and mAP@0.5 by 7.75% and 8.99%, respectively, compared to YOLOv8, demonstrating the effectiveness of this approach in improving traffic light detection accuracy and anti-interference ability.

Key words: traffic lights; target detection; SVDD; YOLOv8; EIoU

摘要： 为提升交通信号灯的检测精度和识别准确率，文章提出基于 YOLOv8 的 SVDD-YOLOv8 目标检测方法，该方法通过整合全局注意力机制（GAM）和支持向量数据描述（SVDD）分 类模块，强化特征捕捉并二次确认目标，同时引入 EIoU 损失函数提高定位精度。对 SVDD 别出的异常区域进行再训练，提升模型性能。实验显示，该方法较 YOLOv8 在检测精度和 mAP@0.5 上分别提升 7.75%和 8.99%，证明了该方法在提高交通信号灯检测精度和抗干扰能 力上的有效性。

关键词: 交通信号灯；目标检测；SVDD；YOLOv8；EIoU