Reducing Depth Measurement Uncertainty in Industrial Robot Stereo Vision Through Error-Aware Disparity Refinement

Stereo vision measurement in industrial robot environments fails on thin structures such as needles and gripper tips, because both traditional and deep-learning-based stereo matching produce boundary errors there. We identify and verify experimentally that ground-truth disparity inflation in public datasets is a systematic error source that pushes networks toward biased boundary estimates. From this, we classify the resulting measurement deviations into four categories: edge errors, intra-segment expansion, small-object loss, and uniform-region distortion. We then propose a geometric-constraint-driven measurement refinement method that corrects each category in turn. Since every correction step is geometric rather than learned, the method is unaffected by ground-truth inflation, a property that conventional post-processing filters do not offer. A GUM-based uncertainty propagation analysis of the measurement model D=f·B/d shows that disparity uncertainty dominates the depth uncertainty budget when f and B are exactly known. Experiments on KITTI 2015, Middlebury, and a custom UE4 synthetic industrial dataset (100 stereo pairs) with nine stereo baselines (seven deep learning and two traditional) show that, on inflation-free ground truth, the refinement imposes a near-zero systematic penalty on deep learning output while clearly improving traditional methods. On KITTI, the predictable metric shift confirms that the method is unaffected by LiDAR ground-truth inflation. On a real industrial robot scene, the refined disparity recovers the gripper tip and needle that the baseline LEAStereo loses. These results position geometric-constraint-driven refinement as an effective, training-data-independent complement to end-to-end stereo matching for precision industrial measurement, within the tested scenes and methods.