A GPU accelerated particle filter based localization using 3D evidential voxel maps 2019-01-0491

In autonomous driving, point cloud from the LIDAR is widely used for localization. Using the 3D point cloud, the localization estimates the pose by matching with an HD map built in advance. For the point cloud based localization, an occupancy voxel map is used as the HD map. The occupancy voxel map divides the environment into three-dimensional discrete space, noted as the voxel. The states of the voxels can be inferred by whether the voxel is occupied by the point cloud or not using an evidence theory. Since the evidence theory provides the additional voxel states such as ‘unknown’ and ‘conflict’, the evidential occupancy voxel map (EOVM) can model the environment with considering the dynamic objects, which makes the EOVM suitable for pose estimation. A Monte-Carlo localization (MCL) is utilized for localization with the EOVM. The MCL estimates the vehicle’s pose with probabilistically spread particles. The particles make the particle filter estimate the pose even if the system model is non-linear. For this reason, the particle filter has the advantage for localization compared with other filtering methods such as Extended Kalman filter, etc. However, there is a computational limitation of the MCL because of the huge amounts of the point cloud, and particles. If the number of particles and the number of points is large, the real-time operation of the localization cannot be guaranteed. In order to overcome the limitation, this paper proposes a multi-core processor based parallel computing framework with optimization techniques to accelerate the computing power. Experiments were performed to evaluate the performance of the localization system in a complex environment, and to compare the computational time between various types of processing units. The experimental results show that the computation time of the proposed parallel particle filter is about ten times faster than particle filter without parallel computing.