计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (5): 123-130.DOI: 10.3778/j.issn.1002-8331.2005-0416

• 模式识别与人工智能 • 上一篇    下一篇

基于改进Cascade RCNN的车辆目标检测方法

李松江,吴宁,王鹏,李海兰   

  1. 长春理工大学 计算机科学技术学院,长春 130022
  • 出版日期:2021-03-01 发布日期:2021-03-02

Vehicle Target Detection Method Based on Improved Cascade RCNN

LI Songjiang, WU Ning, WANG Peng, LI Hailan   

  1. School of Computer Science and Technology, Changchun University of Science and Technology, Changchun 130022, China
  • Online:2021-03-01 Published:2021-03-02

摘要:

针对车辆目标检测过程中小目标及遮挡目标的错检、漏检问题,提出改进Cascade RCNN车辆目标检测方法。使用改进的特征金字塔将浅层信息逐层融入深层网络,增强小目标及遮挡目标特征;引入多支路空洞卷积,减少下采样过程中的特征丢失;将感兴趣区域与上下文信息通过ROI Align统一尺寸后融合,增强目标特征表达。实验结果表明,改进后Cascade RCNN能更好地检测出小目标及遮挡目标,在KITTI和UA-DETRAC数据集上比Cascade RCNN提高了2.2个百分点和2.7个百分点。

关键词: 车辆检测, 小目标, 遮挡目标, Cascade RCNN, ROI Align

Abstract:

Aiming at the problem of misdetection and missed detection of small targets and occlusion targets in the process of vehicle target detection, an improved Cascade RCNN vehicle target detection method is proposed. Firstly, the improved feature pyramid is used to integrate the shallow information into the deep network layer by layer to enhance the small target and occlusion target features. Then the multi-branch dilated convolution is introduced to reduce the feature loss during the downsampled process. Finally, the region of interest and context information through ROI Align unify size and fuse to enhance the target feature expression. Experimental results show that the improved Cascade RCNN can better detect small targets and occlusion targets, which is 2.2 percentage points and 2.7 percentage points  higher than that of Cascade RCNN on the KITTI and UA-DETRAC data sets.

Key words: vehicle detection, small targets, occlusion targets, Cascade RCNN, ROI Align