BIM Informed Visual SLAM for Construction Environments

Problem Statement

Monitoring building construction requires comparing the as-planned design (typically a Building Information Model, or BIM) with the as-built state in real-time. While Visual Simultaneous Localization and Mapping (Visual SLAM) offers a framework for reconstructing environments and localizing sensors, it suffers from significant trajectory drift in construction sites. These environments often feature ambiguous observations and repetitive structures, causing localization errors to accumulate. This drift results in geometrically inaccurate maps, compromising the reliability of as-built vs. as-planned comparisons essential for progress tracking. While LiDAR-based systems have successfully integrated architectural priors to mitigate drift, this integration has not been explicitly explored within visual SLAM frameworks, where sparser depth data and higher noise make wall association more challenging.

Methodology: ivS-Graphs

The authors propose ivS-Graphs, a system that augments an existing RGB-D visual SLAM backbone (vS-Graphs) with structural priors derived from a BIM. The system operates in three primary stages:

1. System Architecture & Inputs:
   The system ingests RGB-D camera data and architectural walls extracted offline from a BIM model. Both detected walls (from the camera) and BIM walls are represented as planar entities defined by a unique identifier, supporting plane parameters (normal and distance), centroid position, length, and thickness.

2. Wall Matching & Initialization:

   • Initial Alignment: The system requires a lightweight human interface to identify correspondences between two perpendicular BIM walls and their detected counterparts. Using these two pairs, the system estimates the transformation between the BIM coordinate frame and the SLAM map frame via a least-squares method. This step aligns the models sufficiently to enable subsequent automated matching.
   • Continuous Matching: As new walls are detected, the system associates them with BIM counterparts using a combined score based on two metrics:
     • Plane-Parameter Distance (PPD): Measures angular and distance residuals between plane equations.
     • Lateral Centroid Distance (LCD): Measures the Euclidean distance between wall centroids projected onto the supporting plane to disambiguate walls with similar plane equations in repetitive layouts.
       Matches are fixed once established and are not re-evaluated in subsequent frames.

3. BIM-Constrained Back-End Optimization:
   The core contribution lies in the back-end factor graph optimization. The system constructs a global graph containing:

   • Keyframes: Camera poses.
   • Map Points: 3D landmarks.
   • Walls: Both detected walls (optimized) and BIM walls (held fixed as the as-planned reference).

   Wall-to-wall factors are introduced as constraints between matched detected walls and fixed BIM walls. The cost function for these factors is weighted by a covariance matrix derived from the matching score; high-confidence matches exert strong constraints, while ambiguous matches are down-weighted. A robust Huber kernel is used to further reduce the influence of outliers caused by partial structures or occlusions. The optimizer minimizes a total cost function comprising visual reprojection errors, pose-to-wall constraints, and the new BIM wall-to-wall factors, thereby correcting keyframe poses globally and bounding drift.

Key Contributions

The paper identifies three main contributions:

1. Novel Integration: The first integration of architectural BIM priors into a visual SLAM framework, enforcing structural consistency between the as-built map and the as-planned design to reduce trajectory drift.
2. Efficient Initialization Strategy: A wall-based initialization and association strategy that requires only two walls as prior information, enabling deployment from the earliest stages of construction without extensive prior mapping.
3. Robustness: A system capable of maintaining mapping accuracy under partially built conditions and geometric discrepancies between the as-planned model and the actual environment.

Experimental Results

The system was validated on 12 real-world sequences across multiple construction sites and office environments, using an RGB-D camera (Intel RealSense D435i) and comparing against state-of-the-art baselines (ORB-SLAM3, BAD-SLAM, DROID-SLAM, and the vS-Graphs backbone).

• Trajectory Accuracy: ivS-Graphs achieved an average Absolute Trajectory Error (ATE) reduction of 25.23% compared to the vS-Graphs baseline. Improvements were most significant in longer trajectories (over 50m), where drift accumulation is highest.
• Mapping Accuracy: The system showed a 7.14% improvement in map accuracy (measured by Root Mean Square Error against a reference point cloud) over the baseline.
• Wall Matching Precision: The system achieved a mean wall-matching precision of 92.8%.
• Robustness:
  • Incomplete Models: When 30% of BIM walls were omitted, the system experienced only a 5.3% increase in ATE, demonstrating resilience to missing structural data.
  • Geometric Deviations: Under severe geometric deviations (simulating construction errors), the system degraded predictably. The authors note that rejecting ambiguous associations (via the Huber kernel) performed better than accepting incorrect matches, preventing catastrophic failure.
• Real-Time Performance: The system operated at an average of 23.3 FPS, meeting the threshold for real-time applications, with the BIM-specific operations (matching and optimization) adding negligible computational overhead.

Significance and Claims

The paper claims that ivS-Graphs successfully addresses the limitation of visual SLAM drift in construction environments by leveraging the semantic and geometric richness of BIM data. Unlike previous approaches that rely on room-level semantics or require pre-existing dense maps, this system operates effectively in partially constructed spaces using only wall features.

The authors emphasize that their approach allows for the creation of drift-bounded as-built maps that remain consistent with the as-planned design, facilitating reliable progress tracking and Augmented Reality (AR) visualization. They acknowledge limitations, such as the requirement for a human to identify the initial two walls for alignment, but argue this is a necessary trade-off to enable operation in single, partially built spaces where full-room localization is impossible. Future work is proposed to integrate additional BIM elements (windows, columns) to further enhance alignment and reduce the need for manual initialization.