PhD Thesis Proposal
Carnegie Mellon University
1:00 pm - 2:00 pm
Depth sensors for robust navigation must measure scenes in darkness, bright light, and in scattering media. Scanning LIDAR devices are the most robust to these conditions, but capture sparse measurements, are slow, and expensive. Consumer depth cameras, on the other hand, are inexpensive and produce dense, high rate depth measurements, but fail in bright ambient light, and are susceptible to effects of global light transport.
Epipolar imaging devices combine the robustness of scanning LIDAR with the speed, sampling density, and potential price-point of consumer depth cameras, but no longer capture images with a global shutter. Due to the line-by-line sampling required by epipolar imaging, these type of depth cameras produce rolling shutter artifacts and are not amenable to traditional localization and mapping methods that pertain to LIDAR and RGB-D sensors.
The goal of this thesis is to develop active illumination depth cameras, sensing methodologies, and navigation algorithms that produce accurate and detailed maps in challenging real-world conditions. This thesis will exploit the unique and robust sensing characteristics of custom-designed depth cameras with novel sensing and mapping algorithms to generate reliable pose and map estimates in challenging environments including those with strong ambient light, complete darkness, smoke, and underwater.
There are three major accomplishments to date. The first is an empirical examination and dataset collection of common robotic sensing technologies deployed in smoke-filled environments. The second is a family of active illumination depth cameras based on epipolar imaging that efficiently capture depth images, are robust to global light transport, and work outdoors in bright sunlight. The third contribution is a prototype active illumination device that only images light from a programmable 3D surface. This device uses triangulation-based range-gating to create steerable curtains of light in the scene that provide inherent obstacle detection and are much less susceptible to scattering media than traditional active imaging techniques.
Proposed research includes developing a continuous-time, spline-based localization and mapping method that overcomes the innate motion warp of rolling shutter epipolar depth cameras to sample the scene and generate dense maps and accurate pose estimates. In addition, a revised hardware prototype of the light curtain device will be developed that has improved frame-rate, range, and reliability in scattering media. These devices and methods will be evaluated in various challenging environments including smoke-filled, underwater, and bright outdoor scenes. The work proposed in this thesis could find applications in a diverse set of fields including mobile robotics, autonomous vehicles, and underwater inspection.
Thesis Committee Members:
William “Red” Whittaker, Co-chair
Srinivasa Narasimhan, Co-chair
Matthew Johnson-Roberson, University of Michigan