The Robust Semantic SLAM System for Texture-Less Underground Parking Lot

Abstract

Automatic valet parking (AVP) is the autonomous driving function that may take the lead in mass production. AVP is usually needed in an underground parking lot, where the light is dim, the parking space is narrow, and the GPS signal is denied. The traditional visual-based simultaneous location and mapping (SLAM) algorithm suffers from localization loss because of inaccurate mapping results. A new robust semantic SLAM system is designed mainly for the dynamic low-texture underground parking lot to solve the problem mentioned. In this system, a 16-channel Lidar is used to help the visual system build an accurate semantic map. Four fisheye cameras mounted at the front, back, left, and right of the vehicle are also used to produce the bird's eye view picture of the vehicle by joint calibration. The vehicle can localize itself and navigate to the target parking lot with the semantic segmented picture and the preobtained semantic map. Based on the experiment result, the proposed AVP-SLAM solution is robust in the underground parking lot.