Overcoming catastrophic forgetting in radar and LiDAR object detection in rain via layer freezing and data augmentation
File(s)radar2025final.pdf (5.44 MB)
Accepted version
Author(s)
Type
Conference Paper
Abstract
Advanced Driver-Assistance Systems (ADAS) use sensors like radar, LiDAR, and cameras for reliable vehicle perception in different weather conditions. While LiDAR and cameras offer high-resolution perception in clear weather, radar excels in adverse conditions such as low light, fog, or rain. Adapting systems trained on clear-weather data to cope with adverse weather often causes catastrophic forgetting, significantly reducing their initial performance after re-training. Unsupervised domain adaptation (UDA) techniques aim to address this but are complex. In this paper, we examine catastrophic forgetting effects on radar and LiDAR, proposing methods to reduce it: model freezing, pre-training with mixed data, and adding simulated data. Our experiments on the well-established RADIATE dataset show these methods improve clear-weather retention and rain detection, with radar showing a 6.59% reduction in forgetting and a 17.19% rain detection gain, and LiDAR a 13.62% reduction in forgetting and 24% improvement with simulations.
Date Acceptance
2025-06-21
Citation
2025 IEEE Radar Conference (RadarConf25)
Publisher
IEEE
Journal / Book Title
2025 IEEE Radar Conference (RadarConf25)
Copyright Statement
Subject to copyright. This paper is embargoed until publication. Once published the author’s accepted manuscript will be made available under a CC-BY License in accordance with Imperial’s Research Publications Open Access policy (www.imperial.ac.uk/oa-policy).
License URL
Source
2025 IEEE Radar Conference (RadarConf25)
Publication Status
Accepted
Start Date
2025-10-04
Finish Date
2025-10-09
Coverage Spatial
Krakow, Poland