Performance evaluation of neural network based integration of vision and motion sensors for vehicular navigation

The positioning accuracy of global navigation satellite systems (GNSS) in dense urban canyon environment significantly deteriorates due to multipath and signal blockage. For this reason, inertial navigation system (INS) is often integrated with GNSS to ensure a reliable navigation solution during such periods of GNSS signal outages. A low-cost navigation solution for land vehicles has been developed by integrating GNSS positioning solution with the measurements from the vehicle motion sensors (accelerometers and gyroscopes). The major drawback of the usage of these inertial sensors is its progressive error accumulation, where the gyroscope drift errors increase gradually, leading to an unusable position estimate, especially in the absence of GNSS updates. Navigation in GNSS-denied environment requires aiding INS with other exteroceptive sensors such as cameras to guarantee the continuity of reliable positioning updates. The estimation of the camera's relative change in position and orientation over time is known as visual odometry (VO). A VO-based multi-sensor integrated navigation system is presented here to surmount the inaccuracy of GNSS in urban scenarios and the drifts of the motion sensors. To enhance the overall system accuracy of the VO-based integrated solution, this paper explores improving the positioning accuracy during GNSS outages by nonlinear modeling of the residual position errors using a neural network. The results show a significant accuracy improvement over relatively long GNSS outages.