Non-Learning Stereo-Aided Depth Completion Under Mis-Projection via Selective Stereo Matching

We propose a non-learning depth completion method for sparse map captured using light detection and ranging (LiDAR) sensor guided by pair of stereo images. Generally, conventional stereo-aided methods have two limiations. (i) They assume the given is accurately aligned to input image, whereas alignment difficult achieve in practice. (ii) limited accuracy long range because estimated pixel disparity. To solve abovementioned limitations, we selective matching (SSM) that searches most appropriate value each image from its neighborly projected LiDAR points based on an energy minimization framework. This selection approach can handle any type mis-projection. Moreover, SSM has advantage terms long-range it directly uses measurement rather than acquired stereo. discrete process; thus, apply variational smoothing with binary anisotropic diffusion tensor (B-ADT) generate continuous while preserving discontinuity across object boundaries. Experimentally, compared previous state-of-the-art completion, proposed reduced mean absolute error (MAE) estimation 0.65 times demonstrated approximately twice more accurate range. under various LiDAR-camera calibration errors, MAE 0.34-0.93 methods.