Parallel Computing in Mobile Robotics for RISE

Abstract

RISE − Risky Intervention and Surveillance Environment is very demanding task for mobile robot where time is crucial. It can be assumed that on-line task execution is very important, therefore the research in parallel computing applied in mobile robotics is needed. Nowadays many researchers are focused on such approaches that uses GPGPU (General Purpose Graphics Processing Unit) for improvement State of The Art (SoA) algorithms. In this chapter three areas of research are shown such as 3D data registration, robot navigation and 3D cloud of points processing for normal vector computation, all are improved by GPGPU computation. The on − line data registration problem is discussed. The approach based on robust KNN k-nearest neighborhood search applied for improvement of ICP algorithm is shown. The path planning parallel approach based on modified diffusion method is shown. On− line 3D cloud of points’ segmentation based on normal vector computation is presented. The set of proposed algorithms where tested on GPGPU NVIDIA CUDA GF 580, the results are satisfying. The improvement of SoA algorithms based on CUDA implementation shows on-line advantages during real task execution. Robust ICP algorithm is needed in mobile robotics applications where data registration has to be performed on−line. New generation of Time of Flight and RGB-D cameras will offer better accuracy and resolution, therefore GPU accelerated data registration algorithms will improve robot navigation, obstacle avoidance and map building. It can be stated that commercial 3D scanners based on rotated lasers offer data acquisition time < 3 seconds, therefore ICP that works in this time will be enough for on-line map building in stop-scan fashion. For this reason robust data registration algorithm based on 3D space decomposition is proposed and the ICP (Iterative Closest Point) approach is chosen as registration method. Algorithm in current version performs matching of two cloud of points up to 512 × 512 = 262144 in 300ms for 30 ICP iterations (NVIDIA GF 580). The proposed solution is efficient since it performs nearest neighbor search using a bucket data structure (sorted using CUDA primitive) and computes the correlation matrix using parallel CUDA all-prefix-sum instruction. The amount of processed points can be increased by implementation on NVIDIA GPU with Compute Capability 2.1. 8