Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Change of Scenery: Unsupervised LiDAR Change Detection for Mobile Robots

Alexander Krawciw, Jordy Sehn, Timothy D. Barfoot

TL;DR

This work tackles unsupervised LiDAR change detection for mobile robots by reframing semantic labeling as per-point Changed vs Consistent classification against a stored map. It introduces RangeNetCD, a range-image CNN that processes map and live-scan pairs under a differentiable loss L = $\mathcal{L}_{\text{cham}} + \lambda_1 \mathcal{L}_{\text{class}} + \lambda_2 \mathcal{L}_{\text{temporal}}$, leveraging Chamfer, class-balance, and temporal terms to learn a fast, unsupervised detector. A retroreflective labeling approach enables rapid ground-truth evaluation, and experiments on forested/off-road data show RangeNetCD substantially outperforms a geometric baseline with mIoU ranging from $67.4\%$ to $82.2\%$, while running at ~10 Hz and integrating into a robot’s autonomy stack for safe navigation. The results demonstrate practical viability for real-time obstacle avoidance and open avenues for future improvements via contrastive encoders and live fine-tuning.

Abstract

This paper presents a fully unsupervised deep change detection approach for mobile robots with 3D LiDAR. In unstructured environments, it is infeasible to define a closed set of semantic classes. Instead, semantic segmentation is reformulated as binary change detection. We develop a neural network, RangeNetCD, that uses an existing point-cloud map and a live LiDAR scan to detect scene changes with respect to the map. Using a novel loss function, existing point-cloud semantic segmentation networks can be trained to perform change detection without any labels or assumptions about local semantics. We demonstrate the performance of this approach on data from challenging terrains; mean intersection over union (mIoU) scores range between 67.4% and 82.2% depending on the amount of environmental structure. This outperforms the geometric baseline used in all experiments. The neural network runs faster than 10Hz and is integrated into a robot's autonomy stack to allow safe navigation around obstacles that intersect the planned path. In addition, a novel method for the rapid automated acquisition of per-point ground-truth labels is described. Covering changed parts of the scene with retroreflective materials and applying a threshold filter to the intensity channel of the LiDAR allows for quantitative evaluation of the change detector.

Change of Scenery: Unsupervised LiDAR Change Detection for Mobile Robots

TL;DR

This work tackles unsupervised LiDAR change detection for mobile robots by reframing semantic labeling as per-point Changed vs Consistent classification against a stored map. It introduces RangeNetCD, a range-image CNN that processes map and live-scan pairs under a differentiable loss L = , leveraging Chamfer, class-balance, and temporal terms to learn a fast, unsupervised detector. A retroreflective labeling approach enables rapid ground-truth evaluation, and experiments on forested/off-road data show RangeNetCD substantially outperforms a geometric baseline with mIoU ranging from to , while running at ~10 Hz and integrating into a robot’s autonomy stack for safe navigation. The results demonstrate practical viability for real-time obstacle avoidance and open avenues for future improvements via contrastive encoders and live fine-tuning.

Abstract

This paper presents a fully unsupervised deep change detection approach for mobile robots with 3D LiDAR. In unstructured environments, it is infeasible to define a closed set of semantic classes. Instead, semantic segmentation is reformulated as binary change detection. We develop a neural network, RangeNetCD, that uses an existing point-cloud map and a live LiDAR scan to detect scene changes with respect to the map. Using a novel loss function, existing point-cloud semantic segmentation networks can be trained to perform change detection without any labels or assumptions about local semantics. We demonstrate the performance of this approach on data from challenging terrains; mean intersection over union (mIoU) scores range between 67.4% and 82.2% depending on the amount of environmental structure. This outperforms the geometric baseline used in all experiments. The neural network runs faster than 10Hz and is integrated into a robot's autonomy stack to allow safe navigation around obstacles that intersect the planned path. In addition, a novel method for the rapid automated acquisition of per-point ground-truth labels is described. Covering changed parts of the scene with retroreflective materials and applying a threshold filter to the intensity channel of the LiDAR allows for quantitative evaluation of the change detector.
Paper Structure (22 sections, 5 equations, 7 figures, 6 tables)

This paper contains 22 sections, 5 equations, 7 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Change Detection
  4. Map Refinement
  5. Teach and Repeat Framework
  6. Methodology
  7. Problem Definition
  8. Chamfer Loss
  9. Class-Balance Loss
  10. Temporal-Consistency Loss
  11. Deep Network Architecture
  12. Range Image Generation
  13. Neural Network Architecture
  14. Experiments
  15. Retroreflective Semantic Evaluation
  16. ...and 7 more sections

Figures (7)

  • Figure 1: The Clearpath Warthog UGV with the Ouster OS-1 LiDAR, driving a previously taught path. A section of the path is now blocked and the change detection algorithm proposed in this paper will be used to allow the robot to safely navigate around the obstruction. The mannequin is wearing the retroreflective suit that is used for dataset generation and evaluation.
  • Figure 2: The dataset trajectories and sample views. In total, the dataset is 8.5 km long with 1.75 km of unique paths.
  • Figure 3: Data flow of the training procedure. Only a single map and live-scan pair are used for inference.
  • Figure 4: A LiDAR scan of two pedestrians coloured by intensity. Left: A pedestrian wearing regular clothes. Right: A pedestrian wearing the reflective suit.
  • Figure 5: Changed IoU vs Distance Traveled Fine-tuning for a map-voxel of 0.3 m and live voxel of 0.05 m. Pre-training improves the performance.
  • ...and 2 more figures