Point Cloud Change Detection With Stereo V-SLAM: Dataset, Metrics and Baseline

Zihan Lin, Jincheng Yu, Lipu Zhou, Xudong Zhang, Jian Wang, Yu WangIEEE RA-L 2022

Introduction

As autonomous vehicles and drones become increasingly integrated into urban environments, their reliance on accurate 3D maps for navigation grows. However, urban landscapes are inherently dynamic; buildings are constructed, structures are demolished, and environments are constantly reshaped. Maintaining up-to-date maps through traditional means is often prohibitively expensive. While "crowdsourcing" map updates using data from consumer-grade vehicles or delivery robots is a promising alternative, it faces significant hurdles in hardware costs and computational complexity. Most existing change detection methods rely on expensive LiDAR sensors, and visual methods like Structure-from-Motion (SfM) are often too slow for real-time applications.

This paper introduces PPCA-VINS, a framework designed to detect changes in 3D point cloud maps using low-cost stereo cameras and Inertial Measurement Units (IMUs). By leveraging prior map data to assist localization and employing efficient change detection algorithms, the authors provide a scalable baseline for visual-based map maintenance.

The PPCA-VINS Framework

The proposed architecture, Prior Point Cloud-Assisted VINS (PPCA-VINS), builds upon the VINS-Fusion framework. Unlike standard Visual-Inertial Odometry (VIO), which estimates motion in a vacuum, PPCA-VINS actively utilizes an existing—though potentially outdated—prior point cloud to refine localization.

Localization and Depth Estimation

The system employs Semi-Global Matching (SGM) to generate semi-dense point clouds from stereo image pairs. To mitigate the noise inherent in single-frame depth estimation, the framework uses temporal filtering, re-projecting depth information across multiple frames.

To align the local data with the global context, the system uses a multi-threaded version of Generalized Iterative Closest Point (GICP). By registering the current local point cloud against the prior map, the system fuses prior geometric information into the pose optimization process. This approach significantly reduces the "drift" common in standard VIO, leading to more stable and accurate trajectories even when the prior map contains outdated sections.

Change Detection Methodology

Once the vehicle's pose is optimized, the system initiates the change detection module. A critical challenge in this process is distinguishing between an actual change in the environment and an area that simply hasn't been observed yet.

Observed Area Construction

To solve this, the authors utilize "Octomaps" and ray-casting to define the camera's frustum—the specific volume of space the camera actually "saw" during its trajectory. This prevents the system from generating false positives in unvisited areas.

Classification of Points

Points within the observed area are classified into two primary categories:

  • New Points: Structures or objects present in the current scan but absent from the prior map.
  • Removed Points: Structures present in the prior map that are no longer detected within the current observed area.

Dataset and Experimental Setup

To evaluate the framework, the authors developed a high-fidelity simulated dataset using Unreal Engine 4 and AirSim. The dataset features three distinct urban environments (Original, S1, and S2), where buildings were systematically added or removed to simulate real-world changes.

The data includes 800x450 stereo pairs at 30Hz and 200Hz IMU data, accompanied by ground truth 6D poses. To test the limits of visual sensors, the authors included challenging scenarios such as mirrors (which complicate depth estimation) and low-light conditions. Additionally, the localization component was validated using the standard EuRoC MAV dataset.

Key Insights and Results

The Virtuous Cycle of Prior Maps

One of the paper's most significant findings is that using an outdated map actually improves current localization. Even if specific buildings are missing or new ones have appeared, the remaining static geometry provides enough constraints to stabilize the trajectory. In testing, PPCA-VINS achieved a Root Mean Square Error (RMSE) of 1.09m on the "S1_01" trajectory, a substantial improvement over the 1.76m RMSE produced by standard VINS-Fusion.

Efficiency and Robustness

The pipeline is designed for computational efficiency. A 170-second trajectory can be processed in approximately 240 seconds on a standard desktop, with the change detection specific tasks requiring less than five minutes. Furthermore, the system demonstrated high recall for detecting new and removed points even when significant noise was introduced to the image and IMU data—conditions under which standard VINS-Fusion often failed.

Sensitivity to Pose Accuracy

The research highlights that change detection is extremely sensitive to localization. Even minor errors in the estimated 6D pose can result in "ghost" points—misaligned data that appears as new objects but is actually just a shifted representation of existing structures like floors or walls. This underscores the necessity of the prior-assisted localization step.

Conclusion and Future Work

PPCA-VINS demonstrates that low-cost stereo-inertial systems can effectively perform 3D change detection when assisted by prior maps. While the system outperforms standard VIO baselines, the authors note that pose error remains the primary bottleneck for precision. Future research may explore alternative map representations, such as semantic or feature-based maps, to further enhance robustness. Additionally, transitioning from high-fidelity simulations to real-world testing will be essential to account for the unpredictable lighting and hardware artifacts found in physical environments.

Summary updated

BibTeX
@article{lin2022point,
  author = {Zihan Lin and Jincheng Yu and Lipu Zhou and Xudong Zhang and Jian Wang and Yu Wang},
  title = {{Point Cloud Change Detection With Stereo V-SLAM: Dataset, Metrics and Baseline}},
  journal = {IEEE Robotics Autom. Lett.},
  year = {2022},
}
Figure