Publication Date

Spring 2026

Degree Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Engineering

Advisor

Jun Liu; Bertin Cordova-Diba; Kaikai Liu

Abstract

Autonomous-driving systems generate large LiDAR point clouds, but compression can damage the sparse object-support structure needed by 3D detectors even when reconstructions appear visually plausible. This thesis asks whether adaptive LiDAR compression can preserve downstream detection better than uniform compression by allocating more fidelity to detector-relevant regions. The main study builds a mask-aware range-image codec with an encoder-decoder bottleneck, an importance head, and an adaptive quantization variant. It compares this adaptive variant package with a confirmed masked uniform baseline under one fixed RangeDet evaluation surface and one fixed KITTI validation subset. Two supporting studies bound the result: a projection-reconstruction PointPillars route and a RENO native point-cloud codec context study. On the controlled main-study surface, the adaptive variant package improves the confirmed uniform baseline by 9.0% at the lower detector threshold and 21.6% at the primary detector threshold, while reducing the in-family rate proxy by 71.1%. The strictest localization threshold improves directionally but remains low, so it is treated as a limitation-aware check. Error analysis shows that the adaptive row recovers more true positives, removes false positives, and improves localization quality in mid-range, far-range, and dense scenes. A control ablation shows that the importance head contributes materially, so the thesis defends the adaptive variant package rather than adaptive quantization alone. The projection-reconstruction route fails confirmatory rerun, and the non-rate-matched RENO adaptive point remains bounded support; broader deployment claims remain future work.

Download

Share

COinS