Efficient 3D LIDAR based loop closing using deep neural network

Loop closure detection in 3D LIDAR data is an essential but challenging problem in SLAM system. It is important to reduce global inconsistency or re-localize the robot that loses the localization, while is difficult for the lack of prior information. We present a semi-handcrafted representation learning method for LIDAR point cloud using siamese convolution neural network, which states the loop closure detection to a similarity modeling problem. With the learned representation, the similarity between two LIDAR scans is transformed as the Euclidean distance between the representations respectively. Based on it, we furthermore establish kd-tree to accelerate the searching of similar scans. To demonstrate the performance and effectiveness of the proposed method, the KITTI dataset is employed for comparison with other LIDAR loop closure detection methods. The result shows that our method can achieve both higher accuracy and efficiency.