Complementary depth integration using convolutional autoencoder

3D spatial recognition is a fundamental technology that supports automatic driving. For example, the processing accuracy of the vehicle depends on the accuracy of depth information around the vehicle body. While methods to geometrically measure the depth of the captured space by applying stereo vision to images taken by multiple cameras become being widely used, it is difficult to measure depth in poorly textured or occluded regions. On the other hand, it becomes possible to estimate depth information in such areas with the advent of estimating depth from monocular images by deep learning. However, if the observation conditions differ between training and estimation, the accuracy of the estimation will decline. This paper proposes a complementary method that integrates both methods by using a convolutional autoencoder.