Precision Analysis of a Variable-Baseline Stereo Vision System for Depth Estimation

Abstract

Stereo vision systems utilise binocular disparity to estimate depth, governed by the trade-off that wide baselines offer better long-range resolution while narrow baselines minimise the near-field blind zone. Most industrial systems employ a fixed baseline, restricting their effective working volume. This paper investigates the depth measurement precision of a custom-built reconfigurable stereo rig to quantify these structural limitations across five baseline configurations. Two Logitech C270 HD cameras were mounted on a 2020-series aluminum V-slot extrusion rail, providing an adjustable baseline from 80mm to 160mm. An independent per-baseline calibration was executed using Zhang’s planar checkerboard method, yielding a stable virtual focal length in the range 915px to 947px across all configurations. The image processing pipeline employed the Semi-Global Block Matching (SGBM) algorithm augmented with a Weighted Least Squares (WLS) disparity filter for subpixel refinement. Experiments measured target distances from 0.5m to 3.0m at 10cm increments, validated against a laser distance meter, producing 390 data points. Results confirm that depth error grows with distance in agreement with triangulation theory. The 80mm baseline achieves a mean relative error (MRE) of 3.33% at 0.6m, while the 160mm configuration suffers a catastrophic 3900.81% MRE at 0.5m owing to SGBM search-space saturation. Conversely, wider baselines (140mm – 160mm) maintain MRE between 2.35% and 4.30% throughout the 1.0m – 3.0m range. These findings provide an empirical framework for dynamic baseline selection in reconfigurable robotic depth-sensing applications.