Vehicle Localization Using Convolutional Neural Networks with IMM-EKF for Automated Vertical Parking

Abstract

This paper proposes a method of vehicle localization using Convolutional Neural Networks (CNN) with Interacting Multiple Model (IMM)-Extended Kalman Filter (EKF) for automated vertical parking. The conventional method for localizing a vehicle in a parking space extracts features from the parking space. It calculates the coordinates of a parking spot. Unlike the conventional methods, CNN provides the pose of the ego-vehicle in this paper. Then, to prevent jittering signals from the CNN, we use a model-based estimator, IMM-EKF, to correct the CNN output. The vehicle state is then corrected using IMM-EKF to prevent jittered estimation results. Although using the IMM-EKF does not noticeably reduce RMS errors in the pose, reductions of the maximum errors are attained up to 50%. From the experiment, the proposed method provides a smooth estimation performance of the vehicle localization compared to another method.