Research on Construction of 3D Map and Accuracy Verification of Self-Localization for Automated Driving of Electric Wheelchair Motoki Hatsuda*, Shoko Oikawa, Toshiya Hirose, Shibaura Institute of Technology, Japan KEYWORDS – Point cloud, Localization, Senior car, Noise removal, Accuracy Research and/or Engineering Questions/Objective – In an automated driving system, self-localization is used to determine where the vehicle is traveling on a map, and these technologies need for the safe driving. Validating the accuracy of self-localization realizes the vehicle safety in automated driving. Therefore, the objectives of this study were to construct a 3D point cloud map and evaluate the accuracy of self-localization. Since the 3D point cloud map contained the noise, this study evaluated the accuracy on self-localization when the noise was removed. We aimed to construct the electric wheelchair with the automated driving system for elderly people to drive smoothly. Methodology – This study conducted three experiments of the distance measurement to landmarks, the first experiment was the stationary measurement, which was measured three positions on six maps, the cart was the stationary condition. The second was the moving measurement, which was measured three positions on a map, the cart was the moving condition. The third was noise reduction effect verification experiment, which was used a map with noise reduction, the cart was the stationary condition. The first and second experiments used the L-shaped fence 2m away from the cart and two pillars 7m away from the cart as landmarks, respectively. This study varied the number of laps and moving velocity as variables to investigate the effect of the number of points in the point cloud map on accuracy. Pedestrians walking on the map become noise, and the data of pedestrians was deleted from the map data using MATLAB. Results – In the first and second experiments, maps with more laps and slower moving velocities (maps with more points in the map) tended to have smaller measurement errors when compared to the real scale distance and the distance recognized by software. Six different maps were used to measure distances, and the range of measurement errors was from 5 mm to 3.5 cm. The map with the removed noise had a smaller error than the map with the noise. In this experiment, two pillars were used as landmarks, and removing the noise reduced the measurement error by 7% and 30%, respectively. Limitations of this study – When measuring distances on the software side, there were errors because each point is too small to manually select the exact location of the target. The distance between the cart and landmarks cannot always be measured on the software. In the second experiment, the cart moved between each position, but stopped to measure the distance. Noise was removed manually, so it is necessary to detect and remove noise in the point cloud map with high accuracy using machine learning. What does the paper offer that is new in the field including in comparison to other work by the authors? – In this study, multiple 3D point cloud maps with different numbers of points were evaluated for the accuracy position of the self-localization. In many studies, noise was removed from the point cloud map and the accuracy was verified. In this study, after constructing a map without noise, landmarks were set as distance measurement objects, and the accuracy of self-localization was verified. In addition, the accuracy of the map was verified by constructing the map for the environment in which the electric wheelchair drives, rather than constructing the map for the automated driving of a passenger vehicle. Conclusions – This study mainly evaluated the accuracy of self-localization, and also constructed maps of the environment in which the electric wheelchair drives and verified the accuracy of maps. The results of distance measurement using a map created with multiple parameters showed an overall trend that the greater the number of points in the map, the smaller the measurement error between the real scale distance and the distance recognized by software. In another experiment, the accuracy of self-position estimation was evaluated with and without noise in the map, and the measurement error was smaller when a map without noise was used.

Mr. Motoki Hatsuda, Student, Shibaura Institute of Technology