Abstract:

Pedestrian sizes in urban road scenes vary widely, and dense crowds can easily cause occlusion problems, increasing the difficulty of pedestrian detection. In order to improve the accuracy and real-time performance of pedestrian detection in urban road areas, and better meet the actual needs of driving scenarios, the paper improves the You Only Look Once（YOLO） method. The original YOLO model is divided into a pedestrian feature extraction stage and a pedestrian coordinate regression stage. The shallow features and deep features are multi-scaled to increase the feature extraction effect of the backbone network, by attention mechanism and GIoU loss, adding spatial channel enhancement module after feature fusion to improve recognition ability of occluded targets. Using Cross YOLO layer and adjusting the width of the network in combination with the size of the pedestrian has speeded up the model calculation speed. The verification experiments are performed on the Caltech pedestrian benchmark dataset. The results show that compared with the original YOLO and other popular methods, the YOLO-Person model has a lower false detection rate for small targets and occluded targets, and it is faster and has certain practical applications. value.

Key words: You Only Look Once（YOLO）, pedestrian detection, feature enhancement

摘要：

城市道路场景下的行人目标尺寸变化大，并且人群密集容易引起遮挡问题，增加了行人检测难度。为了提高城市道路区域行人检测的准确性和实时性，更好应对驾驶场景的实际需求，对You Only Look Once（YOLO）方法进行改进。原YOLO模型分为行人特征提取阶段和行人坐标回归阶段，将浅层特征与深层特征多尺度融合，增加骨架网络的特征提取效果；添加注意力机制，在特征融合后加入空间通道增强模块，并且将GIoU损失引入网络训练过程，提高对遮挡目标的识别能力；结合行人尺寸，提出CrossYOLO层对网络宽度进行调整，加快了模型运算速度。在Caltech行人基准数据集下进行验证实验，结果表明YOLO-Person模型与原YOLO以及其他流行方法相比，对小目标和遮挡目标误检率更低，并且速度更快，具有一定的实际应用价值。

关键词: YOLO方法, 行人检测, 特征增强