Neural Network Adaptation of the Kalman Filter for Odometry Fusion

In navigation systems for unmanned vehicles, an important task is fusion of the pose estimations (odometry and localization) obtained from different sensors: cameras, LiDARs, wheel encoders, inertial measurement modules, etc. To solve this task, it is necessary to know the covariance matrices for each of the odometry sources, which characterize the prediction accuracy of the corresponding pose. In this paper we propose a neural network adaptation of noise covariances in Kalman filter for odometry fusion task. Instead of specifying process and measurement noise covariances manually, we use optimization technique and neural network on input time series data to automatically predict values of these covariance matrices. Our approach fuses the vehicle 3D position and orientation obtained using visual and LiDAR-based methods for simultaneously localization and mapping. The experiments were conducted on a dataset from unmanned ground robot Clearpath Husky. Comparing to Kalman filter without adaptation, our method is more precise. We made a software implementation of the proposed approach based on PyTorch deep learning framework.