Low-Cost SPAD Sensing for Non-Line-Of-Sight Tracking, Material Classification and Depth Imaging


CLARA CALLENBERG, University of Bonn, Germany
ZHENG SHI, Princeton University, USA
FELIX HEIDE, Princeton University, USA
MATTHIAS B. HULLIN, University of Bonn, Germany

Link: https://light.princeton.edu/publication/cheapspad/

1. Background

  • Time-correlated imaging, or the recording of the optical response of a scene to transient illumination, allows to analyze the temporal dimension of light transport, a feature that is not accessible in pure intensity imaging
  • Time-correlated optical measurements have established themselves as a valuable source of information
  • The approaches available for recording time-correlated measurements are rich and varied, but most require bulky and expensive hardware and are too fragile to be used outside of lab settings
  • A notable exception is the emerging technology of single-photon avalanche diodes (SPADs)
  • Single-photon avalanche diodes
  • Time-of-flight (ToF), transient and depth imaging
  • Non-line-of-sight (NLOS) tracking
  • Material classification

3. Main Work


  • Propose to use an off-the-shelf sensor evaluation kit as a lowcost alternative to high-end SPAD sensors, and equip the board with a custom firmware to output raw photon count histograms
  • Introduce hardware add-ons such as collimating optics and galvanometer scanners to meet the needs of a selection of key applications for time-resolved imaging. Further propose reconstruction pipelines based on inverse filtering, deep learning, and other computational sensing paradigms that are capable of handling the low-resolution time-tagged measurements produced by our system
  • Validate the proposed platform for some of the most iconic application modes of time-resolved imaging, namely non-line-of-sight object tracking, material classification, and depth imaging
  • Propose cost-neutral feature additions to the sensor hardware that would greatly improve their interfacing to external hardware, and their suitability as a general-purpose sensing platform for time-resolved light transport

3.1. Low-Cost SPAD System

  • VL53L1X time-of-flight sensor module by STMicroelectronics
  • The 12-pin package, priced around USD 3 for large volumes, has a footprint of 15mm^2 and integrates a 940nm light source and a 16X16 SPAD array sensor with a field of view of 27 degrees imaged by a miniature lens
  • Use additional optical equipments including glasses and galvanometer scanners for increased flexibility of the system

3.2. Material Classification

  • When placing the sensor right onto the surface of a material, the infrared light from the VL53L1X light source penetrates the material, is scattered inside, and part of it is reflected back to the SPAD sensor
  • Depending on the structure of the material, the signal measured by the sensor can vary temporally and spatially
  • By training a neural network, characteristics of different materials can be learned and they can later be distinguished by holding the sensor to an object

3.3. Tracking Objects “Around the Corner”


  • VL53L1X can be used to track an object “around the corner” by illuminating a wall facing the hidden area and recording the echoing light signal that is reflected from the target object
  • Train a neural network to recognize the target position from the SPAD data of four measurements on the wall

3.4. Depth Imaging


  • The VL53L1X can yield a spatially resolved transient image by scanning all possible 4×4 ROIs on the 16×16 pixel sensor, which yields a 13×13 pixel measurement
  • Use additional glasses and galvanometer scanners to avoid the substantial blur due to the overlapping ROIs and the poor optical quality of the imaging lens
  • Depth maps are calculated in two different ways:
    • Calculate the given pixel’s depth as the weighted mean of the captured histogram. This way we achieve even smooth depth gradients and sub-bin accuracy in the depth estimation
    • Compute the depth by fitting Gaussian functions to the histogram of each pixel, which yields sharper and more reliable results at the cost of a longer runtime