According to the problem that the multi-scale vehicle objects in traffic surveillance video are difficult to detect and the overlapping objects are prone to missed detection, an improved vehicle object detection method based on YOLOv3 was proposed. In order to extract feature more efficiently, we first use the inverted residuals technique to improve the convolutional layer of YOLOv3. To solve the multi-scale vehicle object detection problem, three spatial pyramid pooling(SPP) modules are added before each YOLO layer to obtain multi-scale information. In order to cope with the overlapping of vehicles in traffic videos, soft non maximum suppression (Soft-NMS) is used to replace non maximum suppression (NMS), thereby reducing the missing of predicted boxes due to vehicle overlaps. Our experiment results in the Car dataset and the KITTI dataset confirm that the proposed method achieves good detection results for vehicle objects of various scales in various scenes. Our method can meet the needs of practical applications better.

针对交通监控视频中多尺度目标难以检测，重叠目标容易漏检的问题，提出了一种改进的基于YOLOv3的目标检测方法。为了更有效地提取特征，我们首先使用反向残差技术来改进YOLOv3的卷积层。为了解决多尺度车辆目标检测问题，在每个YOLO层之前添加了三个空间金字塔池（SPP）模块以获得多尺度信息。为了应对交通视频中车辆的重叠，使用了非最大抑制（Soft-NMS）代替了非最大抑制（NMS），从而减少了由于车辆重叠而导致的预测框丢失。我们在Car数据集和KITTI数据集中的实验结果证实，该方法对于各种场景中各种比例的车辆对象均能获得良好的检测结果。我们的方法可以更好地满足实际应用的需求。