A Vision-based Forward Driving Vehicle Velocity Estimation Algorithm for Autonomous Vehicles

In this paper, we proposed a vehicle velocity estimation algorithm when it is in front of autonomous car based on the only vision system. In order for autonomous cars to drive safely, they need to recognize their surrounding drive environment using various sensors. Among them, the stereo camera can acquire not only a normal image but also an image showing a depth map. Therefore, in this paper, the speed of the target vehicle was estimated by using only the stereo camera without any other sensor. Also, to recognize the target vehicle, CNN based semantic segmentation model with high speed and accuracy was used. Unlike object detection models, the semantic segmentation model can acquire only the object shape, so there is a lot of depth information that can be obtained from the depth image. Through this, more accurate distance information between own car and target vehicle was extracted. Lastly, the relative speed between vehicles was estimated by using the estimated distance change and the time difference between frames. The superiority of the proposed algorithm performance was compared with the vehicle speed using a motor encoder. And it was confirmed that the speed of the target vehicle was accurately estimated using only the camera.