MACHINE LEARNING FRAMEWORK FOR VISUAL TRACKING

摘要

A method of analyzing autonomous vehicle data comprising recording a video of a vehicle environment utilizing one or more vehicle cameras, identifying corner points of objects in the video, identifying a forward-tracked location of one or more corner points in each frame from an earlier frame to a later frame of the recorded video played in forward, identifying a reverse-tracked location of one or more corner points in each frame from the later frame to the earlier frame of the recorded video played in reverse, comparing the forward-tracked location of the earlier frame and reverse-tracked location of the later frame, and adjusting a descriptor defining characteristics of one or more pixels of the corner point in response the comparison indicating an error rate exceeding a threshold.