A Light-weight stereo matching network for an embedded vision system

In the embedded computer vision field, the stereo vision matching using two cameras can identify the distance of the surrounding environment at low prices. Recently, CNN based artificial intelligence algorithms have been showing good results with high accuracy. However, huge network size and slow computation prevent them used in embedded edge systems. In this paper, a CNN algorithm to use reduced parameters to find a accurate disparity map of stereoscopic images based on Siamese network is proposed. The proposed method performs network training through GPU-based parallel operation and produces better results than the existing one using fewer parameters.