一种多尺度YOLOv3的道路场景目标检测算法

A multi-scale YOLOv3 detection algorithm of road scene object

针对在自然交通场景中道路不同种类目标的边界框大小差异巨大,现有实时算法YOLOv3无法很好地平衡大、小目标的检测精度等问题,重新设计了YOLOv3目标检测算法的特征融合模块,进行多尺度特征拼接,对检测模块进行改进设计,新增2个面向小目标的特征输出模块,得到一种新的具有5个检测尺度的道路目标多尺度检测方法YOLOv3_5d.结果表明:改进后的YOLOv3_5d算法在通用自动驾驶数据集BDD100K上的检测平均精度为0.5809,相较于原始YOLOv3的检测平均精度提高了0.0820,检测速度为45.4帧·s^(-1),满足实时性要求.

In natural traffic scene,the bounding box sizes of different road targets vary greatly.The existing real-time object detection algorithm YOLOv3 can not balance the detection accuracy of large and small targets and has poor performance in the task.To solve the problems,the feature fusion module of YOLOv3 target detection algorithm was redesigned to realize the multi-scale feature stitching.The detection module was improved by adding two extra feature output modules for small targets,and a new multi-scale detection method of YOLOv3_5d for road targets was obtained with 5 detection scales.The experimental results show that the average precision of the improved YOLOv3_5d algorithm is 0.5809 on BDD100K dataset,which is 0.0820 higher than that of original YOLOv3.The running speed is 45.4 frames·s^(-1),which can meet the real-time requirement.