Robust 3D Perception Under Limited Supervision and Sparse Sensing

Dissertation

Three-dimensional perception enables autonomous vehicles, robots, and intelligent environments to understand objects, motion, geometry, and human activity. However, its practical deployment is often constrained by expensive manual annotation, sparse and incomplete geometric measurements, noisy sensor data, and computational limitations. These challenges affect critical applications such as autonomous navigation, human–robot interaction, rehabilitation, assisted living, and privacy-aware monitoring.

This dissertation addresses these challenges through four contributions. It introduces a learned affinity network for LiDAR-based 3D multi-object tracking and a self-supervised framework that reduces the need for manually labeled temporal identities. It also presents a sparse-to-dense reconstruction method that combines images with limited LiDAR measurements to improve 3D object detection. Finally, it develops an efficient mmWave-radar framework for low-latency, privacy-aware 3D human pose estimation. Together, these contributions advance reliable and practical 3D perception under limited supervision, sparse sensing, noisy observations, and real-time deployment constraints.