A Dead Reckoning Method Based on Neural Network Optimized Kalman Filter

Inertial Measurement Unit (IMU) plays a key role in real-time trajectory estimation. The commonly used low-cost IMU is affected by the deviation and noise, and when the acceleration is double integrated to obtain the displacement, the error will increase indefinitely. Traditional state estimation methods rely on an accurate system model, and the estimation accuracy of low-cost IMU is low. In the motion system, the main reason that affects the model accuracy and estimation accuracy is the fluctuation of lateral and vertical velocities. Firstly, this paper uses a convolutional neural network to learn the relationship between sensor data and lateral and vertical velocities and their variances and constructs a noise adapter to obtain pseudo-measured values and variances. Secondly, the weights of the adapter output and the state estimation value are distributed in the Kalman filter, and the variance of the adapter output is used as the measurement covariance to update the Kalman gain, to obtain the optimal estimation value. Finally, according to the estimated speed value and the initial position, the attitude and position are calculated. Training and verification on the KITTI data set have achieved good results.