Publication Type
PhD Dissertation
Version
publishedVersion
Publication Date
3-2026
Abstract
Spatial computing has been gaining popularity primarily due to advancements in sensing and reasoning capabilities over three-dimensional (3D) spaces. Light Detection and Ranging (LiDAR) technology is a key enabler of this progress, with LiDAR sensors now widely integrated into devices ranging from mobile phones to domestic robots due to their growing commercial availability and affordability. While LiDARs provide highly accurate 3D reconstructions of the physical world, applications that rely on them continue to face several performance drawbacks, including high energy consumption, mutual interference, and fundamental hardware limitations that impose undesirable tradeoffs between range, frame rate, and resolution. This thesis investigates and validates a set of system-level innovations for core LiDAR operations to address these challenges, enabling more efficient, robust, and scalable LiDAR-driven spatial computing across diverse devices and environments.
First, this thesis presents D2SR, a middleware for scanning LiDARs designed to mitigate mutual interference in environments where multiple LiDAR-equipped pervasive devices operate independently in close proximity with overlapping fields of view. D2SR reduces the impact of interference on LiDAR depth estimation through a lightweight three-stage process consisting of decentralized detection, mitigation, and recovery. It detects interference using a machine learning-based classifier, minimizes interference by time-shifting the duty cycle of individual LiDARs without requiring explicit communication between LiDAR-equipped devices, and recovers corrupted depth frames by employing generative AI models to synthesize missing or inaccurate depth estimates.
Next, this thesis explores NeuroLiDAR, a sensor fusion framework that integrates LiDAR sensing with neuromorphic event cameras to enable adaptive and high-frame-rate depth estimation. LiDAR performance is often constrained by hardware limitations that require trade-offs between range, spatial resolution, and frame rate. Many LiDAR systems operate at low frame rates, typically between 5 and 10 Hz, favouring long-range sensing over responsiveness to rapid scene changes. NeuroLiDAR addresses this limitation through an adaptive depth sensing framework that fuses infrequent and intermittent sparse LiDAR data with high-frequency inputs from neuromorphic event cameras. It incorporates two key components: (a) event-based keyframe detection, which activates LiDAR scanning infrequently, and (b) event-guided depth extrapolation, which uses the event sensor stream to synthesize depth values in between such scans. These innovations allow the sensing rate to adjust dynamically in response to scene activity while preserving LiDAR accuracy.
Finally, this thesis introduces MuLES, a novel multistatic LiDAR architecture that achieves low-power, infrastructure-assisted depth sensing and provides built-in trajectory estimation capability for mobile pervasive devices. MuLES offloads the power-intensive emitter functionality to the environment’s infrastructure, where its energy-intensive requirements can be met. The mobile devices are equipped with a passive receiver that comprises a set of photodiodes, significantly reducing the energy demand on battery-operated devices. MuLES replaces the standard time-of-flight (ToF) calculation with an angle-of-arrival (AoA)-based method to estimate egocentric point clouds from the receiver’s perspective, without requiring any synchronization with the emitter. The emitter additionally encodes its pulses to differentiate among them. Such encoding allows the receiver’s position to be localized within its own coordinate system without relying on computationally intensive processes such as SLAM.
Together, these system-level breakthroughs advance the efficiency and adaptability of LiDAR sensing by pervasive devices, and enable easier, robust, and scalable deployment of future spatial computing applications.
Keywords
Mutual Interference, Crosstalk, High-frame-rate LiDAR, LiDAR, Multistatic LiDAR
Degree Awarded
PhD in Computer Science
Discipline
Computer Sciences
Supervisor(s)
MISRA, Archan
First Page
1
Last Page
146
Publisher
Singapore Management University
City or Country
Singapore
Citation
KANATTA GAMAGE, Ramesh Darshana Rathnayake.
Refining LiDAR-driven depth perception for pervasive spatial computing. (2026). 1-146.
Available at: https://ink.library.smu.edu.sg/etd_coll/845
Copyright Owner and License
Author
Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial-No Derivative Works 4.0 International License.