IMU and LIDAR Odometry Based on CNN and RNN Self-Supervised Learning and Attention Mechanism

摘要

Robots use their own sensors to obtain environmental perception information to solve their own understanding of the environment, such as SLAM(Simultaneous localization and mapping). This paper uses self-supervised deep learning to solve the odometry problem of robots in SLAM by modeling the fusion of IMU measurements and lidar measurements. In this paper, we uses a neural network to fuse the IMU measurements of two lidar data frames with lidar data. Fusing more deterministic a priori information can effectively reduce the trajectory drift for filter-based or optimized-based odometry. This paper also adopts this idea to increase the length of the processing sequence by using the feature that LSTM can process multiple frames of lidar data simultaneously. The global positional pose is also obtained using the LSTM’s ability to remember prior information. The learning-based approach can simplify the modeling difficulty and get better odometry results, but the supervised learning-based method needs to get the accurate position and pose true value. In many cases, obtaining accurate positional pose truth is very expensive and difficult to obtain. Therefore, it is reasonable and feasible to form a selfsupervised approach to the network’s output by fully using the training data’s characteristics. Based on this, in this paper, we propose a network structure that utilizes CNN and RNN and incorporates IMU measurement information. We propose neural network odometry using IMU and lidar data fusion. Using the feature that the LSTM can change the sequence length, we design a processing method similar to the sliding window algorithm to suppress drift, taking full advantage of the inspiration of the sliding window algorithm. Introducing a self-supervised adversarial mechanism based on point cloud segmentation. It makes the self-supervised process more robust and fast convergence.