Fast Motion Estimation Method for Self-driving Video

As auto-driving cars become increasingly dependent on computer vision technology, the storage and real-time transmission problems of massive videos from self-driving cars follow. Efficient coding technology helps to save storage space and increase transmission speed. However, the current encoding time is still too long to meet the speed requirements of video storage and transmission in self-driving cars. Under this circumstance, this paper proposed a fast motion estimation method for self- driving scene videos called DFME. It is an algorithm based on the regulars in driving scene videos. Such regular makes it possible to simplify the model prediction, motion vector prediction, and motion vector search process. It has been proved in experiments on HEVC that the method DFME can reduce 30% encoding time on average with less than 1% PSNR lost for self-driving scene videos in Kitti data set.