基于改进YOLOv3的车辆检测算法

交通场景下的车辆检测问题存在小目标多、目标遮挡严重等情况,鉴于此,提出一种基于改进YOLOv3的车辆检测算法.由于小目标仅包含较少的像素,特征不明显,算法在空间金字塔结构中融入软池化操作,搭建Soft-SPP结构将多重感受野融合,通过软池化操作最大程度地保留细节,有效提取小目标特征;引入坐标注意力机制,在调整每个通道特征分配权重的同时能够捕捉具有精确位置信息的远程依赖关系;提出一种新的损失函数KIoU Loss作为边界框损失函数,同时考虑边界框的关键点与长宽比使之回归更加准确.实验结果表明,改进后的算法在自动驾驶KITTI数据集上平均精度达到94.69%,相比原始YOLOv3算法精度提升4.13%,且检测速度仅下降3.16 frame·s^-1,在保持检测速度的情况下能够明显提升检测精度.

Vehicle detection algorithm based on improved YOLOv3

Aiming at the problems of vehicle detection in the traffic scene such as a large number of small targets and severe target occlusion, a single-stage target detection algorithm based on the improved YOLOv3 is proposed. Since the small target only contains fewer pixels and features are not obvious, this algorithm builds a Soft-SPP structure based on the idea of spatial pyramid pooling, which integrates multiple receptive fields and adopts soft-pooling operation to retain details to the maximum extent and avoids information loss. The coordinate attention mechanism is introduced to capture the remote dependence with accurate location information. and adjust the weight assigned to each channel feature to make the network better learn important information. A loss function KIoU Loss based on key points and aspect ratio is proposed as the boundary box loss function, which makes the boundary box regression more accurate. The experimental results show that the mAP of the improved algorithm on the autopilot KITTI data set is 94.69%, which is 4.13% higher than that of the original YOLOv3 algorithm, and the detection speed is only reduced by 3.16 frame•s^{-1