Possible approach you can implement
- For your video input, grab single frames
- Calculate keypoints in every frame
- Match keypoints in subsequent frames and calculate the optical flow
- Relate the optical flow values to actual speed values (will need some kind of calibration for the car I guess)
