A Comparison of Motion Priors for EKF-SLAM in Autonomous Race Cars

Abstract: Simultaneous Localization and Mapping (SLAM) is one of the fundamental problems to solve for any autonomous vehicle or robot. The SLAM problem is for an agent to incrementally build a map of its surroundings while keeping track of its location within the map, using various sensors. The goal of this thesis is to demonstrate the differences and limitations of different odometry methods when used as the prior motion estimate for the common SLAM algorithm EKF-SLAM. Inspired by autonomous racing, the algorithms were evaluated for especially difficult driving scenarios such as high velocities and while the car was skidding. Three different odometry algorithms that all rely on different sensors were implemented; The feature based stereo visual odometry algorithm Libviso2, the lidar odometry algorithm Range Flow-based 2D Odometry (RF2O), and wheeled odometry fused with measurements of the vehicles angular velocity from a gyroscope. The different algorithms were evaluated separately on real data that was gathered by running a modified RC car equipped with the necessary sensors around different racing track configurations. The car was driven for different levels of aggressiveness, where more aggressive driving implies a higher velocity and skidding. The SLAM estimate of the vehicle position and cone locations were evaluated in terms of mean absolute error (MAE) and computational time, for all motion priors separately on each track. The results show that Libviso2 provides an accurate prior motion estimate with consistent performance over all test cases. RF2O and the wheeled odometry approach could in some of the cases provide a prior motion estimate that was sufficient for accurate SLAM performance, but performed poorly for other cases.