This repository manages the tasks and documentation for the Master's Thesis: "Lite-BEV: Lightweight LiDAR-Radar BEV Fusion for Robust 3D Object Detection in Adverse Weather".
Detailed documentation, work packages, and meeting notes can be found in the project Wiki: 👉 Go to Wiki (or use the GitLab Wiki tab)
This thesis investigates multi-modal sensor fusion methods combining LiDAR and 4D Radar for 3D object detection in autonomous driving, specifically focusing on performance under adverse weather conditions (rain, fog).
To address the limitations of LiDAR in poor visibility, this work proposes Lite-BEV, a lightweight Bird’s-Eye View (BEV) fusion framework. The approach avoids computationally expensive Transformer-based cross-attention and instead leverages two core mechanisms:
- Doppler-Masked Geometric Distillation: A training strategy where a 4D Radar encoder learns LiDAR-quality geometric features, specifically for dynamic objects identified via Radar's radial velocity.
- Aleatoric Uncertainty-Driven Gating: An inference-time fusion gate that dynamically switches between LiDAR and Radar based on estimated LiDAR reliability (e.g., spatial variance or point density drops in fog/rain).
The method is built upon the OpenPCDet framework, using PointPillars as a baseline, and is evaluated on the MANtruckscenes dataset.
docs/workpackages/: Detailed technical descriptions of each work package.wiki/: Project documentation and progress logs (synced as a submodule)..gitlab/issue_templates/: Templates for creating work-package related issues.
To initialize the wiki submodule:
git submodule update --init --recursive