Tracking moving objects in a video sequence by the neural network trained for motion vectors

This paper demonstrated the use of the neural network to track the motion of moving objects recorded in a sequence of video images. The motion vectors measured by the exhaustive search algorithm at the grids of a mesh set on an image were used as part of inputs to train the neural network. The coordinates of such pixels that exhibited a significant displacement relative to the overall image displacement caused by camera blurring, etc. were input along with pixel color data. Given the motion vector of an arbitrary pixel and color, the neural network determines if the pixel belongs to the moving object or not. The work intends to apply a neural network to estimate motion vectors to analyze the structure of the moving object and recognize the geometrical shape or structures from motion information, by means of the inverse problem referred to as "structure from motion" (SfM) problem.