Deep Visual Odometry for Robot Navigation
We implemented DeepVO, a deep visual odometry model (paper 1, paper 2), to estimate a robot’s motion from a sequence of camera images, and integrated it with the ROS navigation stack.
To test its effectiveness we added Gaussian noise to the wheel odometry and compared navigation performance with and without DeepVO in the localization loop. In the end DeepVO was added as a measurement model to the EKF localization algorithm.
Result
The red dot is the target destination.