Vision-based Vehicle Ego Velocity Estimation

Abstract

Vehicle ego velocity estimation is an interesting topic of research, as a reliable and accurate estimation method is crucial in vehicle motion control. At the same time, smart vehicles are often equipped with multiple sensors, such as cameras, radar, and so on. By incorporating the vision-based velocity method, we can increase the system redundancy in case of other sensor failures. In this thesis, we proposed a method that can predict accurate forward velocity as well as leftward velocity. In our experiments, the best performance is 0.526(km/h) MAE and 0.751(km/h) RMSE for forward velocity estimation, and 0.171(km/h) MAE and 0.250(km/h) RMSE for leftward velocity estimation. Moreover, to make our method more reliable in practical use, we also performed uncertainty estimation on the model with the best performance, which makes our method more applicable.