曹文冠 黄孝慈 舒方林 孙昊 刘景锋



摘 要: 无人车在狭窄、没有GPS信号的地下停车场中行驶,精确的定位能力非常重要。传统的基于视觉的定位方法由于无纹理的区域、重复的结构和外观的变化而导致跟踪丢失。本文利用视觉语义信息构建停车场地图并对车辆进行定位。语义特征包括经常出现在停车场中的路标、停车线、减速带等,与传统特征相比,这些语义特征无纹理区域和外观的变化具有长期稳定和鲁棒性。文中采用单目相机对外界环境进行感知。在惯性测量单元(IMU)和车轮编码器的协助下,提出一种生成全局语义地图的方法。此地图用于厘米级别上车辆定位。分析了该系统的准确性和召回率,并与现有的方法进行了比较。最后,通过自动停车的应用验证了该系统的实用性。
关键词: 视觉语义信息;语义地图;定位;单目相机
文章编号: 2095-2163(2021)03-0205-05 中图分类号:TN929.51 文献标志码:A
【Abstract】For unmanned vehicles to drive in narrow underground parking lots without GPS signals, accurate positioning capabilities are very important. Traditional vision-based localization methods lead to loss of tracking due to untextured regions, repetitive structures and changes in appearance. This paper uses visual semantic information to construct a parking lot map and locate vehicles. Semantic features include road signs, parking lines, speed bumps, etc, which often appear in parking lots. Compared with traditional features, these semantic features have long-term stability and robustness for their textureless areas and appearance changes. The research uses monocular cameras to perceive the external environment. With the assistance of IMU (Inertial Measurement Unit) and wheel encoders, a method to generate a global semantic map is proposed. This map is used for vehicle positioning on the centimeter level. The accuracy and recall rate of the system are analyzed and compared with existing methods. Finally, the practicability of the system is verified through the application of automatic parking.
【Key words】 visual semantic information; semantic map; location; monocular camera
0 引 言
近年來,城市发展过程中,停车问题已经成为了一个亟待解决的研究热点。作为解决这一问题的有效方法,自动泊车技术逐渐受到多方关注。同时研究可知,该技术也是自动驾驶技术的重要组成部分[1]。自动泊车系统的主要功能包括:高精度定位、车位识别、轨迹规划与路径跟踪控制。系统运行时,车载环境感知传感器(如视觉传感器、激光雷达、超声波雷达等)开始工作,感知车辆周边的车位情况,并确定车位的相对位置信息;然后,由系统控制单元判定车位是否可……