Kamijo Lab. covers a wide range of topics related to autonomous vehicles, such as object detection and scene understanding by deep learning techniques, and self-localization and digital mapping. In this article, we would like to introduce research on self-localization and digital mapping.

Currently, point cloud data is generally used for research and demonstration of Autonomous Driving. Each point consists of three-dimensional coordinates detected by LiDAR or stereo vision sensors. Buildings and roadside facilities such as traffic signals are configured with point cloud data in the digital map. Autonomous vehicles estimate their position by matching detected points to the point clouds in the digital map as the reference.

Generally, SLAM can provide point clouds with high accuracy of local alignment that is relatively accurate in the digital map. On the other hand, SLAM suffers from lower accuracy of global alignment that is absolutely accurate in the digital map. For example, a digital map might have around five meters of error in global alignment when SLAM is applied to 1km of driving data. This problem can be solved by matching point cloud data to an airborne image which has accurate absolute coordinates [1]. Popular digital maps with purposes other than autonomous driving consist of vector data. Thus, the point cloud map is considered unique for the demonstration of autonomous driving [2], and we aim to solve the problems of those gaps between the point cloud map and existing popular digital maps [1][2].

The error of self-localization increases inherently when the environment has fewer or featureless buildings and structures [3]. An error of more than 1m is not unusual in such circumstances. In addition to such static factors, dynamic factors such as occlusion of the LiDAR beam by tall buses and trucks strongly affect the accuracy of self-localization [4]. We aim to formulate a quantitative method to evaluate the localization error affected by both the static and dynamic environments [3][4].

[1] Mahdi Javanmardi, Ehsan Javanmardi, Yanlei Gu, Shunsuke Kamijo, “Automatic Calibration of 3D Mobile Laser Scanning using Aerial Surveillance Data for Precise Urban Mapping”, IEEE Intelligent Vehicles Symposium (IV2017), Redondo Beach, California USA, June, 2017.
[2] Ehsan Javanmardi, Mahdi Javanmardi, Yanlei Gu, Shunsuke Kamijo, “Autonomous Vehicle Self-Localization Based on Multilayer 2D Vector Map and Multi-channel LiDAR”, IEEE Intelligent Vehicles Symposium (IV2017), Redondo Beach, California USA, June, 2017.
[3] Ehsan Javanmardi, Mahdi Javanmardi, Yanlei Gu, Shunsuke Kamijo, “Adaptive Resolution Refinement of NDT Map Based on Localization Error Modeled by Map Factors”, 21st International Conference on Intelligent Transportation Systems (ITSC2018), Maui, HI, USA, Nov. 2018.
[4] Yuki Endo, Ehsan Javanmardi, Yanlei Gu, Shunsuke Kamijo, “Evaluation of High Definition Map-Based Self-Localization Against Occlusions in Urban Area”, to appear in IEEE Intelligent Vehicles Symposium (IV2021), Nagoya, July 2021.

Text: Shunsuke Kamijo (Associate professor)
English proofreading: David Buist (Senior project specialist)