Fisheye Stereo Vision: Depth and Range Error

This study derives analytical expressions for the depth and range error of fisheye stereo vision systems as a function of object distance, specifically accounting for accuracy at large angles.

💡 Research Summary

The paper presents a rigorous analytical treatment of depth and range errors in fisheye stereo vision systems, explicitly accounting for large incident angles that are typical in wide‑field‑of‑view (WFOV) applications. Starting from the well‑known pinhole camera model, the authors derive the classic depth‑error relationship ΔZ = Z²·Δd/(f·B) and then convert it to range error ΔR = Z²·Δd/(f·B′) where the effective baseline B′ = B·cosθ shortens as the viewing angle θ moves off the optical axis. This formulation predicts a 1/cosθ increase in error for peripheral points.

The core contribution lies in extending the analysis to the equidistant (f‑θ) fisheye model, where the radial image coordinate follows r = f·θ. By expressing disparity as d = f·arctan(B/(2Z)) and differentiating with respect to Z, the authors obtain a depth‑error expression ΔZ = Z²·Δd/(f·B)·√(1 + tan²θ). Consequently, the range error becomes ΔR = Z²·Δd/(f·B′)·√(1 + tan²θ). The √(1 + tan²θ) factor (equivalently 1 + tan²θ under the square‑root) grows much faster than the pinhole’s 1/cosθ term, indicating that fisheye optics suffer a pronounced degradation of angular resolution (instantaneous field‑of‑view, iFOV) toward the sensor periphery.

To illustrate the practical impact, the authors consider a 4K (3840 × 2160) fisheye camera with 2.1 µm pixel pitch and a full 180° horizontal field of view. Using the equidistant projection, the effective focal length is calculated as f ≈ 1222.3 pixels. With a 1 m baseline, a disparity error of Δd = 0.2 pixels, and a target depth Z = 10 m, the derived model predicts range errors below 4 cm for incidence angles within ±30°. Beyond this angular window, the error escalates according to the 1 + tan²θ term, whereas a comparable pinhole system would only experience the milder 1/cosθ scaling.

The discussion emphasizes that the primary limitation of fisheye stereo is not geometric baseline foreshortening but the loss of angular precision at large θ, which directly amplifies disparity uncertainty. Mitigation strategies therefore focus on increasing the baseline (which is now feasible thanks to recent real‑time auto‑calibration algorithms such as NODAR’s Hammerhead SDK) and on maintaining robust extrinsic calibration under environmental stresses (vibration, wind, temperature).

Limitations of the study include the assumption of a constant disparity error across the image, neglect of residual lens‑distortion errors, sensor noise, illumination variations, and dynamic scene effects. Future work is suggested to incorporate these non‑idealities, validate the analytical predictions with extensive empirical data, and explore adaptive disparity‑error models that reflect pixel‑wise confidence.

In conclusion, the paper delivers closed‑form expressions for depth and range errors in both pinhole and fisheye stereo configurations, quantifies the accelerated error growth inherent to fisheye optics, and demonstrates that with a wide baseline and modern auto‑calibration, fisheye stereo can achieve centimeter‑level range accuracy at moderate distances. These insights are directly relevant to autonomous navigation, robotics, and large‑scale surveillance where WFOV depth perception is essential.