Introduction
Recovering dense 3D models from monocular video sequences is a fundamental challenge in computer vision, particularly for applications like railway corridor mapping. The primary bottleneck is the "correspondence problem"—the computational effort required to match points across successive frames. While traditional methods search along epipolar lines, doing so exhaustively for every pixel or region is prohibitively slow. This paper introduces Temporal Search space Reduction (TSR), an unsupervised learning framework designed to optimize this process by leveraging the repetitive visual nature of corridor environments.
Methodology
The proposed system operates in two main stages: dense planar recovery and the TSR efficiency layer.
Planar Quadrilateral Recovery
Instead of tracking individual pixels, the system divides the image into a 64x64 grid of quadrilaterals. Each cell is treated as a flat plane. The algorithm searches for the 3D orientation (normal) and distance (offset) that minimizes the Sum of Absolute Differences (SAD) when the quadrilateral is reprojected into the subsequent frame. This approach allows for a dense reconstruction that is more computationally manageable than pixel-wise methods.
Temporal Search space Reduction (TSR)
The TSR layer acts as a "visual-spatial memory." For each quadrilateral, the system extracts 15 statistical features—including mean, standard deviation, skewness, kurtosis, and energy—across RGB channels.
As the video is processed in reverse chronological order, the system learns the relationship between these 2D textures and their 3D positions.
If a new frame contains a texture similar to one previously encountered (e.g., track ballast), the system uses stored Gaussian distributions to predict a narrow search range. If the feature is novel, an exhaustive search is performed, and the result is added to the model for future use.
Results and Discussion
The TSR framework was evaluated using a custom dataset of high-definition railway footage. The results demonstrate a significant leap in efficiency without sacrificing reconstruction quality.
- Efficiency: TSR reduced the required focal offset checks by 94%, dropping from an average of 324.73 to just 19.62 per quadrilateral.
- Speed: With a processing time of 1.29 seconds, TSR outperformed established methods like PTAM (9.77s) and MonoSLAM (13.03s).
- Accuracy: Interestingly, restricting the search space improved accuracy by reducing "false positive" matches. The mean error was 0.99 meters, compared to 1.13 meters for exhaustive search and 4.09 meters for SIFT-based methods.
Conclusion
By treating 2D appearance as a predictor for 3D location, the TSR framework effectively "remembers" the geometry of structured environments. This unsupervised approach allows for dense 3D recovery at a fraction of the usual computational cost, making it highly suitable for real-time corridor mapping. Future work will explore the generalizability of this visual memory to more chaotic urban settings.
Summary updated