Trajectory Based Particle Filter: Asynchronous Observation Fusion for Autonomous Driving Localization

• Published in: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC) pp. 114 - 121
• Main Authors: Jonchery, Sylvain, Revilloud, Marc, Ouerhani, Yousri
• Format: Conference Proceeding
• Language: English
• Published: IEEE 19-09-2021
• Subjects: Location awareness; Monte Carlo Localization; Monte Carlo methods; Particle filters; Roads; Sensor Fusion; Sensors; Trajectory; Urban areas
• DOI: 10.1109/ITSC48978.2021.9564793

Estimating the pose of a vehicle is an essential function for an autonomous vehicle. Numerous methods exist to tackle this problem, but they are often specialised for one particular type of road, as they only use one type of landmarks. This paper therefore proposes to used different types of landmark detector into a single coherent framework to deal with all these scenarios. To this end, we used three complementary types of landmarks, lane markers, dense and reliable in motorways and suburban areas, road signs, sparse but reliable, and building walls, dense in cities. A Particle Filter is then used to relocalize a vehicle in a map using these landmarks, a conventional GPS and vehicle odometry. One of particularity of the proposed filter is its ability to handle heterogeneous data from various sensors asynchronously. This approach has been experimentally tested on real data on a mixed scenario of 15 km containing different types of roads from highway to city center roads. Results are encouraging and show that our approach is able to fuse different types of landmarks in the same framework, with a root mean square error (RMSE) around 1 meter.