Building enterance recognition from street-level LiDAR and video images

点击翻译按钮获取中文摘要

Every day, Google Maps helps countless people in cities around the world find buildings they are looking for, but once they get there, they are usually on their own to find the actual entrance to get in. Building entrance locations have particular meaning for autonomous driving, and routing algorithms in map searches where the results can bring users to the exact entry. This is quite useful, especially for a large building block, where current routing algorithms cannot provide accurate entry information due to the lack of entrance location information in the maps. To add entrance information to the maps, manual operation is current practice which is a labor intensive and tedious task. The proposed research is automatic recognition of building entrance from street-level video images and LiDAR data. Building entrances vary greatly in terms of shape and color. They are often partially occluded by pedestrians and other objects from street level data, which makes automatic recognition a very challenging task. The proposed research will develop a novel sensor fusion approach for building entrance detection that takes advantages of video images and LiDAR, which use a graphical model to formulate the joint probability of the candidate detection, image evidence and scene information. We expect that our entrance recognition method will be generic to, and robust enough for, a wide variety of entrances, in terms of shape, color, texture, occlusions, illumination, scales, and viewpoints. The extracted building entrance information can be added to a map database to improve map accuracy and thereby enhance map search capability. Eventually, will make it easier for individuals and eventually autonomous cars to find the front door. The automated process will also significantly improve productivity of map generation.