Occlusion Area Removal in Binocular 3D Reconstruction of Train Running Parts

The state-of-the-art approaches of stereo matching rely on trained CNNs. However, there are not proposed effective methods in the current end-to-end deep learning stereo matching network to solve the problem that a pair of rectified stereo images are matched in the occlusions and background areas, which leading to erroneous reconstruction of 3D point clouds in these areas. In this paper, a softmax module to the back-end of the stereo matching network is added to calculate the confidence of the disparity map. Moreover, the occlusion part and the background part of the disparity map is removed according to the characteristic that the correct area and the occlusion area have a morphological difference on the waveform of the confidence response. Finally, a train running parts dataset is generated to prove our method. This work realizes the rapid three-dimensional reconstruction and measurement of the bottom part of the motor car under the single image acquisition, and providing a feasible solution for removing the occlusion and background parts in stereo matching.