Computational 3D Imaging with Position Sensors

Jeremy Klotz, Mohit Gupta, Aswin C. SankaranarayananICCV 2023 · DOI

Abstract

Traditional 3D laser scanning systems are often limited by the high bandwidth and power requirements of CMOS-based image sensors. This paper introduces a novel 3D scanning architecture that replaces pixelated cameras with Position Sensing Diodes (PSDs)—low-power, non-pixelated sensors that directly output the centroid of incident light. To overcome the PSD's inherent sensitivity to global illumination (interreflections and scattering), the authors propose a "Computational PSD" system. By integrating a Spatial Light Modulator (SLM) and a robust regression-based algorithm, the system can isolate direct light from global interference, enabling high-resolution 3D reconstruction with significantly reduced computational overhead.

Introduction

Standard 3D laser scanners track a laser spot across a scene using CMOS cameras. This approach is inherently inefficient: the sensor captures millions of pixels of data simply to identify the coordinates of a single bright point. This results in high power consumption and significant memory requirements, which are detrimental for mobile or robotic applications.

Position Sensing Diodes (PSDs) offer a compelling alternative. As non-pixelated sensors, they function as analog computers, outputting a single (x, y) coordinate representing the center of mass (centroid) of all light hitting the sensor surface. While incredibly fast and efficient, PSDs traditionally fail in complex environments where light bounces off multiple surfaces (global illumination). Because a PSD cannot distinguish between the direct laser spot and secondary reflections, the resulting centroid is biased, leading to massive errors in 3D shape estimation.

Schematic and equivalent circuit of a Position Sensing Diode (PSD).

Methods

The proposed Computational PSD system addresses the global illumination challenge through a combination of hardware modulation and statistical estimation. The setup consists of a laser projector, a PSD sensor, and an SLM placed within the sensor's optical path.

The SLM acts as a programmable spatial mask. By displaying various high-frequency patterns, such as checkerboards, the system selectively blocks different portions of the scene from reaching the PSD. Rather than relying on a single measurement, the system captures multiple data points under different masks.

The core innovation lies in a robust regression algorithm. The authors derive a linear model where the "direct" light centroid is determined by the relationship between the changes in measured voltages across different masks. Unlike simpler "min-max" separation techniques, this regression approach utilizes all captured data, providing superior robustness to noise and environmental interference.

Key Results

The authors evaluated their prototype using challenging real-world objects, including V-grooves (playing cards) and wax candles, which are notorious for interreflections and subsurface scattering.

  • Accuracy in Complex Scenes: In a V-groove test case, standard PSD measurements resulted in a plane-fitting error of 1598µm. The proposed global suppression method reduced this error to 320µm.
  • Precision: The system achieved a centroid resolution of 5µm at 150mW laser power, demonstrating high-fidelity geometric reconstruction.
  • Data Efficiency: The system successfully reconstructs complex shapes, such as plaster busts, without ever needing to process a traditional 2D pixel array, drastically reducing the data footprint.
Comparison of 3D reconstruction with and without global suppression on a playing card scene.

Discussion and Future Work

One of the most significant advantages of this system is its defocus invariance. Because the PSD tracks the centroid, symmetric blur kernels do not shift the measured position, allowing for sharp 3D scans even when the optical image is out of focus.

Currently, the system's primary limitation is scan time, which is constrained by the refresh rate of the SLM and galvo mirrors. However, the authors suggest that future iterations using MEMS laser projectors and higher-bandwidth PSDs could enable real-time, low-power 3D scanning. This technology holds particular promise for Shortwave Infrared (SWIR) imaging, where high-resolution CMOS sensors remain prohibitively expensive, offering a cost-effective path toward advanced sensing in mobile devices and autonomous robotics.

Summary updated

BibTeX
@inproceedings{klotz2023computational,
  author = {Jeremy Klotz and Mohit Gupta and Aswin C. Sankaranarayanan},
  title = {{Computational 3D Imaging with Position Sensors}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2023},
  doi = {10.1109/iccv51070.2023.00746},
}

Loading related content...

Figure