Chamelion: Reliable Change Detection for Long-Term LiDAR Mapping in Transient Environments

TL;DR

Chamelion tackles reliable online low-dynamic change detection for long-term LiDAR mapping in transient environments by combining a composition-based single-session augmentation with a 4D sparse convolutional backbone and a dual-head network that jointly predicts changes and cross-visibility confidence. The framework enables occlusion-aware change classification and a probabilistic, confidence-driven map update to maintain map consistency across sessions. Key contributions include a novel pseudo-label generation strategy, a 4D backbone with a dual-head architecture, and a Bayes-filter-like map update that robustly fuses multi-session observations. The approach demonstrates strong generalization across real-world construction and indoor environments and achieves efficient update behavior on both desktop GPUs and embedded platforms.

Abstract

Online change detection is crucial for mobile robots to efficiently navigate through dynamic environments. Detecting changes in transient settings, such as active construction sites or frequently reconfigured indoor spaces, is particularly challenging due to frequent occlusions and spatiotemporal variations. Existing approaches often struggle to detect changes and fail to update the map across different observations. To address these limitations, we propose a dual-head network designed for online change detection and long-term map maintenance. A key difficulty in this task is the collection and alignment of real-world data, as manually registering structural differences over time is both labor-intensive and often impractical. To overcome this, we develop a data augmentation strategy that synthesizes structural changes by importing elements from different scenes, enabling effective model training without the need for extensive ground-truth annotations. Experiments conducted at real-world construction sites and in indoor office environments demonstrate that our approach generalizes well across diverse scenarios, achieving efficient and accurate map updates.\resubmit{Our source code and additional material are available at: https://chamelion-pages.github.io/.

Figures (9)

• Figure 1: Our method detects low dynamic (LD) changes in real-time between the prior map and the current scan. Here, red points represent negative changes (NC) that have disappeared from the map, while blue points indicate positive changes (PC) that have newly appeared in the current scan.
• Figure 2: Our proposed composition-based augmentation for the pseudo-label generation. Static objects are extracted using multi-object tracking-based segmentation, where $O^k = \{o^k_j \mid j = {0, \dots, T}\}$ denotes snapshots of the $k$-th object. These objects are pasted onto arbitrary locations in single-session scans ($\mathcal{S}_{0:t}$) and the map ($\mathcal{M^S}$) to form a pseudo multi-session dataset, $\tilde{\mathcal{S}}_{0:{T}}$ and $\tilde{\mathcal{M}}$. Here, $o_\mathrm{pc}$ is inserted into a scan, and $O_\mathrm{nc}$ into the map.
• Figure 3: Our dual-head architecture for change detection. (a) First, we extract features from the input using a 4D convolutional neural network. These features are then fed into two separate heads: one for (b) change classification and the other for (c) cross-visibility confidence estimation. When updating the map, we only trust the class output in areas with high cross-visibility confidence, where both the map and scan are visible.
• Figure 4: Our approach handles occlusion using cross-visibility confidence. Gray points denote static regions, and red points indicate negative change predictions by the class head. (a) Occluded areas (black dashed) caused by walls hinder accurate prediction. (b) The confidence head estimates visibility scores. (c) Class predictions are applied only to visible regions, reducing errors in occluded areas.
• Figure 5: Qualitative comparison of different change detection methods on our custom dataset. The first column represents the prior map, and the second column corresponds to the current scan data. The third to seventh columns visualize the results from other SOTA methods and our proposed approach, overlaying the predictions onto both the map and the scan data. Green, red, and blue points correspond to true changes, false changes, and false statics, respectively. The fewer red and blue points, the better the result.