Learning End-to-End Inertial-Wheel Odometry for Vehicle Ego-Motion Estimation

We present an end-to-end learning architecture for wheel speeds and inertial fusion for real-time ego-motion estimation. In contrast to the previous work, our method utilizes the neural network to learn the nonlinear vehicle model and reduce the modeling error in motion propagation. The network inputs are the wheel speeds and inertial measurements while the outputs are the 2D pose increments in vehicle frames. To enhance the learning efficiency of the network, we propose a data enhancement strategy and employ a large amount of data for training. A diversity of practical scenario tests have shown that our proposed method outperforms traditional methods in terms of accuracy and robustness, achieving an average translation error of less than 0.4m per 100m, which meets the practical application requirements and the trained model has been deployed as a front-end module for other systems on an autonomous vehicle.