结合深度学习与双目视觉的多目标障碍物追踪测距研究  

Multi-target obstacle tracking and ranging based on deep learning and binocular vision

作　　者：赵伟 武帅琦 张意林 王世辉 ZHAO Wei;WU Shuaiqi;ZHANG Yilin;WANG Shihui(School of Vehicle and Traffic Engineering,Henan University of Science and Technology,Luoyang henan 471000,China)

机构地区：[1]河南科技大学车辆与交通工程学院,河南洛阳471000

出　　处：《激光杂志》2023年第10期57-64,共8页Laser Journal

基　　金：河南省科技攻关项目(No.202102210278);河南科技大学研究生创新基金项目(No.CXJJ-2021-CY08)。

摘　　要：针对现有无人驾驶车辆环境视觉识别对于多目标障碍物测距存在鲁棒性较差和误差较大的问题,提出了一种双目视觉与深度学习相结合的多目标障碍物实时追踪测距方法。通过将SGBM立体匹配算法进行改进,添加可自动调节窗口大小的数层次快速均值滤波消除视差图的噪声干扰,再经过开运算的形态学处理方法进行视差空洞填充,然后与YOLOv3提取出的障碍物像素点进行坐标对应,得到障碍物距离值。实验结果表明,所提出的方法能够对不同类型的障碍物进行视距测量,且距离30 m范围内测量相对误差保持在3%以内,平均处理速度为28帧每秒,具有较好的实际意义。The existing environmental vision recognition of unmanned vehicles has the problems of poor robustness and large error for multi-target obstacle location.A multi-target obstacle real-time tracking and location method combining binocular vision and deep learning was proposed.The SGBM stereo matching algorithm was improved,and the noise interference of the parallax map was eliminated by adding the digital-level fast mean filtering which could automatically adjust the window size.Then,the parallax hole was filled by the morphology processing method of open operation,and the obstacle distance value was obtained by corresponding coordinates with the obstacle pixels extracted by YOLOv3.The experimental results show that the proposed method can measure the visual distance of different types of obstacles,and the relative error of measurement within 30 m distance is kept within 3%,and the average processing speed is 28 frames per second,which has good practical value.

关 键 词：双目测距 多目标追踪 深度学习 SGBM算法 YOLOv3目标检测 

分 类 号：TN209[电子电信—物理电子学]