High Accuracy Real-Time 6D SLAM with Feature Extraction Using a Neural Network

抄録

We present a method for three-dimensional (3D), point cloud-based, six-dimensional (6D) simultaneous localization and mapping (SLAM) with high accuracy and very low computational cost. The two key points of our high accuracy and real-time SLAM process are feature extraction and scan matching optimization. For the 3D laser-based SLAM using the iterative closest point algorithm, we consider correct corresponding point pair searching for achieving high accuracy. Therefore, we propose extracting feature points from 3D point clouds for correct corresponding point pair searching. To extract features such as edges and corners in real-time, we propose the use of a trained neural network (NN). The NN used in our feature extraction scheme is a simple backpropagation (BP) NN with two hidden layers, which allows building a real-time system for 6D SLAM. To optimize the scan matching, we propose the use of particle swam optimization (PSO) and the extracted feature points. The PSO increases the accuracy of the estimated position by matching the most features with a global map stitched with all features. Compared with the state-of-art methods, the proposed method achieved the best performance for the KITTI Odometry Benchmark.