A deep learning approach to state estimation from videos

Abstract

Kalman lters have been commonly used for estimating the state of a vehicle from a video. Multi-State Constraint Kalman Filter (MSCKF) is an EKF-based state estimator that uses feature measurements for pose estimation of a vehicle. These models require a lot of hands-on engineering time to dene the measurement functions. We propose a data-driven approach by training deep neural networks on high-dimensional navigation image data generated from a simulation. We describe a CNN model that robustly learns reliable features from the input and gives promising results to model temporal data. We show that a deep learning approach can be a replacement for the MSCKF model for estimating the velocity of a moving vehicle.