Abstract
This paper presents a novel generative framework for modeling local visual features as a function of a robot's pose. By moving away from explicit 3D geometric maps, the proposed system utilizes Radial Basis Function (RBF) networks to learn the appearance, position, and scale of Scale-Invariant Feature Transform (SIFT) features across different viewpoints. This approach enables robust global robot localization and the synthesis of "virtual" views of an environment. Experimental results in two indoor scenes demonstrate mean localization errors as low as 7.7 cm, highlighting the effectiveness of modeling implicit geometry for navigation in complex environments.
Introduction
Traditional robot localization typically relies on either explicit 3D geometric models or global image features. While 3D models provide precision, they are often difficult to construct in "messy" real-world environments characterized by specularities and occlusions. Conversely, global features like PCA are computationally efficient but brittle under varying illumination or dynamic obstacles. This research addresses the gap by modeling local image features—specifically SIFT—within a "Visual Map" framework. The goal is to allow a robot to predict expected observations from arbitrary poses without requiring a perfect 3D reconstruction.
Methodology
The proposed system follows a four-stage pipeline to construct a generative Visual Map:
- Feature Extraction: The SIFT algorithm is employed to detect local peaks in a difference-of-Gaussian pyramid, providing features invariant to scale and rotation.
- Feature Tracking: Features are matched across training images with known poses using a ratio test to ensure distinctiveness.
- Generative Modeling: For each tracked feature, an RBF network learns the mapping from camera pose $(x, y)$ to the feature's appearance, image coordinates, and scale. This allows for continuous interpolation of feature properties across the environment.
- Probabilistic Inference: The models are evaluated via leave-one-out cross-validation to estimate covariance matrices, which represent the "trustworthiness" of each feature during localization.
A key innovation is the use of mixture models for localization. Instead of a simple joint likelihood, which can be derailed by a single outlier, the system sums probabilities. This ensures that the pose estimate remains robust even when some feature matches are incorrect.
Experimental Results
The framework was validated in two indoor environments: Scene I (200x200 cm) and Scene II (300x600 cm). Ground truth was established using high-precision laser tracking systems.
In Scene I, the system achieved a mean localization error of 7.7 cm. In the larger Scene II, the error was 18.4 cm. The generative nature of the model was demonstrated by "rendering" virtual views from novel viewpoints, effectively reconstructing the visual essence of the scene from the learned RBF models.
Analysis of the error distribution revealed that errors were more pronounced in the "looming" direction (along the optical axis) compared to lateral movements, as visual changes are less distinct when moving directly toward or away from a scene.
Conclusion and Future Work
This work demonstrates that implicit geometric modeling via generative RBF networks provides a stable and device-agnostic solution for robot localization. By focusing on local invariant features, the system overcomes many limitations of traditional mapping techniques. Future research will aim to incorporate camera heading ($\theta$) into the pose space and explore online learning methods to build maps dynamically during exploration without pre-existing ground truth.
Summary updated