2019/10/08 by V. Vaquero, Vaquero, Victor, Kai Fischer +7
Engineering · Computer Science · #Robotics and Sensor-Based Localization #Advanced Neural Network Applications #Autonomous Vehicle Technology and Safety
paper · pdf · doi:10.48550/arxiv.1910.03336
Localization and Mapping is an essential component to enable Autonomous\nVehicles navigation, and requires an accuracy exceeding that of commercial\nGPS-based systems. Current odometry and mapping algorithms are able to provide\nthis accurate information. However, the lack of robustness of these algorithms\nagainst dynamic obstacles and environmental changes, even for short time\nperiods, forces the generation of new maps on every session without taking\nadvantage of previously obtained ones. In this paper we propose the use of a\ndeep learning architecture to segment movable objects from 3D LiDAR point\nclouds in order to obtain longer-lasting 3D maps. This will in turn allow for\nbetter, faster and more accurate re-localization and trajectoy estimation on\nsubsequent days. We show the effectiveness of our approach in a very dynamic\nand cluttered scenario, a supermarket parking lot. For that, we record several\nsequences on different days and compare localization errors with and without\nour movable objects segmentation method. Results show that we are able to\naccurately re-locate over a filtered map, consistently reducing trajectory\nerrors between an average of 35.1% with respect to a non-filtered map version\nand of 47.9% with respect to a standalone map created on the current session.\n