UEM-CNN: Enhanced Stereo Matching for Unstructured Environments with Dataset Filtering and Novel Error Metrics

Abstract:
We aim at the passive 3D reconstruction of unstructured environments to analyze terrain navigability, detect objects, and avoid obstacles in mobile off-road robotics. For this purpose we present three patch-based depth reconstruction networks for unstructured environments denoted UEM-CNNbase, UEM-CNN9, and UEM-CNN19. The KITTI 2012 dataset is used for training, validation, and testing. We present filtering for erroneous training data which optimizes network performance. Novel error metrics are introduced which assess the disparity estimation results for their usability in the environmental perception for robotics. We achieve a three pixel error, a maximum deviation of three pixels from the ground truth disparity value, of up to 14.26% averaged over all images in KITTI 2012. On the five unstructured images in KITTI 2012 our UEM-CNN still achieves a remarkable three pixel error of up to 17.01 %. This emphasizes the strong performance of UEM-CNN considering the challenges of depth estimation in completely unstructured environments.