NeuroSense AI: Confidence-Aware Stereo Depth Estimation — An Offline Error Characterization for Assistive Perception
Year: 2026
Type: preprint
Abstract: Reliable spatial sensing is a prerequisite for non-invasive assistive navigation systems. This preliminary study evaluates classical stereo matching and left–right (LR) consistency as components of a proposed confidence-aware depth perception pipeline, NeuroSense AI, using Middlebury stereo data. On the chess1 scene at 50% image scale, the baseline attained 86.31% recall and 95.38% precision for identifying surfaces within 1.5 m. A strict LR consistency filter increased precision to 96.19% but reduced recall to 69.87% and increased the fraction of unassessed pixels from 26.49% to 42.03%. A three-state confidence partition at a 1.0-pixel LR tolerance yielded 57.97% VALID, 15.54% UNCERTAIN and 26.49% INVALID ground-truth pixels; disparity mean absolute error (MAE) was 4.1764 px for VALID versus 13.9023 px for UNCERTAIN pixels. On the separate Motorcycle scene, the full-scene StereoSGBM depth MAE was 0.0338 m over the 81.81% of reference-valid pixels that also had a valid prediction, and one 3D candidate cluster had 93.36% of its projected voxel centroids inside a manually refined object mask. These findings suggest that LR consistency identifies less reliable estimates, but treating rejected estimates as absent can suppress relevant near-surface detections. A separate caution/unknown alert policy is proposed but not yet evaluated. All results concern offline, pixel-level analysis and do not establish pedestrian navigation safety.
Author: Sobirova, Ikboloy Ikrom Kizi
Language: English
Topic: Advanced Vision and Imaging
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7234770447
