Abstract
Markerless Augmented Reality (AR) systems often suffer from visual instability, or "jitter," due to noise in camera pose estimation. This paper proposes a novel post-processing technique using Support Vector Regression (SVR) to stabilize virtual augmentations. By dynamically learning motion patterns from a sliding window of frames, the SVR filter effectively reduces high-frequency oscillations and spurious pose jumps without requiring a predefined motion model.
Introduction
In markerless AR, the system calculates the camera's pose by tracking natural features in the environment. While more flexible than marker-based approaches, this method is prone to two significant issues: pose jitter and spurious estimations. Jitter manifests as high-frequency oscillations of virtual objects, even when the camera is stationary, while spurious estimations cause jarring "jumps" in the object's position. Traditional solutions like Kalman filters often struggle with the unpredictable motion of handheld devices. This work introduces a "plug-and-play" SVR-based regression module to address these challenges.
Methodology
The proposed system inserts an $\epsilon$-Support Vector Regression module between the pose estimation and rendering engines.
SVR Configuration
The system utilizes $\epsilon$-SVR with a Gaussian Radial Basis Function (RBF) kernel. To avoid mathematical singularities, pose estimation is decomposed into seven independent SVRs: four for a unit quaternion (rotation) and three for the translation vector ($x, y, z$). The SVR is retrained on a small sliding window of recent frames to predict the smoothed pose for the current frame.
Latency Mitigation
To prevent lag during rapid camera movements, the authors implemented a "quick move detection" mechanism. By monitoring the centroid of extracted features, the system can detect sudden shifts, bypass the SVR filter to maintain responsiveness, and reset the training window.
Experimental Evaluation
The method was evaluated using a 1,870-frame video sequence of an aeronautical maintenance task. The sequence is characterized by low-texture surfaces and rapid camera motion, presenting a significant challenge for feature-based tracking. A high-quality reference baseline was established using SIFT features on full-resolution images with image doubling.
Results and Discussion
The SVR approach was compared against raw SURF-based pose estimation and a standard Kalman filter. Quantitative results are summarized in the table below.
| Method | RMSE (mm) | Max Error (mm) |
|---|---|---|
| SURF (Raw) | 9.39 | 94.10 |
| SURF + SVR | 14.42 | 57.32 |
| SURF + Kalman | 18.99 | 63.53 |
The results indicate that while any smoothing filter introduces a slight increase in Root Mean Square Error (RMSE) due to inherent delay, the SVR filter significantly reduces the Maximum Error compared to both raw estimation and the Kalman filter. This reduction in peak error is critical for maintaining the "illusion of mixture" in AR by eliminating the most distracting visual artifacts.
Conclusion
The SVR-based jitter reduction technique provides a robust, model-free solution for stabilizing markerless AR. By dynamically adapting to motion patterns, it outperforms traditional Kalman filters in suppressing high-frequency noise and spurious jumps. Future research will explore optimizing the computational cost of per-frame retraining and refining parameter selection for diverse environments.
Summary updated