Abstract: The quality of pedestrian images collected by autonomous vehicle mounted cameras in low illumination traffic scenes is poor due to insufficient lighting and complex environments, and subsequent detection algorithms are difficult to ensure sufficient detection accuracy. Therefore, in response to the problem of poor pedestrian detection performance in low illumination traffic scenes, this paper proposes a pedestrian detection algorithm based on improved YOLOv4-Tiny. First of all, the output of 8 times down sampling feature map is increased for the backbone network, and the deep semantic information and shallow semantic information are fused from bottom to top to enhance the detection ability for small targets. At the same time, the attention mechanism module is introduced before the fusion of different feature maps, making the network pay more attention to key feature information. Secondly, SPP-Net is used to improve the Receptive field and robustness of the network. Using K-means clustering algorithm to generate a new prior box for pedestrian targets, replacing traditional non maximum suppression methods with Soft-NMS method. The improved network model is labeled YOLO-IPD, and experiments have shown that the YOLO-IPD model proposed in the article performs well on a self built dataset.

Key words: Pedestrian detection; Low illumination; YOLOv4-Tiny; Attention mechanism; Deep learning

摘要： 无人驾驶汽车车载相机在低照度交通场景下由于光照不足、环境复杂导致采集的行人 图像质量差，后续检测算法难以保障足够的检测精度。因此，针对低照度交通场景下行人检 测效果不好的问题，文章提出一种基于改进 YOLOv4-Tiny 的行人检测算法。首先，对骨干网 络增加了 8 倍下采样特征图输出，并自下而上的融合深层语义信息和浅层细节信息，以增强 对小目标的检测能力，同时在不同特征图融合之前引入注意力机制模块，使网络更加关注重 点特征信息。其次，使用 SPP-Net 提高网络的感受野和鲁棒性。利用 K-means 聚类算法对行 人目标生成新的先验框，用 Soft-NMS 方法替换掉传统的非极大值抑制方法。改进后的网络模 型记为 YOLO-IPD，实验表明文章提出的 YOLO-IPD 模型在自建数据集上效果良好。

关键词: 行人检测；低照度；YOLOv4-Tiny；注意力机制；深度学习