Abstract
Visual Simultaneous Localization and Mapping (SLAM) is critical for autonomous navigation, yet traditional feature-based methods like ORB-SLAM2 suffer from high computational overhead due to per-frame descriptor extraction. We present FastORB-SLAM, a lightweight system that eliminates this bottleneck. By employing a two-stage, descriptor-independent keypoint matching method, FastORB-SLAM achieves nearly double the tracking speed of ORB-SLAM2 while maintaining state-of-the-art accuracy, particularly in low-texture environments.
Introduction
Visual SLAM systems are the primary sensory mechanism for autonomous robots to map environments and localize themselves. "Indirect" methods, which rely on extracting keypoints and calculating mathematical descriptors for matching, are highly accurate but computationally expensive. A significant portion of these descriptors is used only once for frame-to-frame tracking and then discarded, representing a major source of "descriptor waste." FastORB-SLAM addresses this by decoupling tracking from descriptor computation, reserving the latter only for long-term map maintenance.
System Architecture
FastORB-SLAM builds upon the proven architecture of ORB-SLAM2, consisting of three main threads: Tracking, Local Mapping, and Loop Closure. The primary innovation lies in the Tracking thread, which replaces standard descriptor-based matching with a more efficient coarse-to-fine approach.
Coarse-to-Fine Keypoint Matching
The system utilizes a novel two-stage matching process to establish correspondences between frames without computing descriptors for every point:
- Coarse Matching via Motion Modeling: Instead of an exhaustive search, the system employs a Uniform Acceleration Motion (UAM) model to predict keypoint locations in subsequent frames. This provides a superior "initial guess" compared to constant velocity models. Sparse optical flow is then used on an eight-level image pyramid to find matches based on grayscale intensity.
- Fine Refinement and Outlier Removal: To ensure robustness, two geometric filters are applied. Grid-based motion statistics verify motion smoothness across groups of points, and a RANSAC-based fundamental matrix check ensures matches adhere to epipolar geometry constraints.
Descriptors are computed only for "keyframes"—frames selected for long-term storage in the map—enabling efficient loop closure and relocalization without the per-frame overhead.
Experimental Results
The system was evaluated on the TUM RGB-D and ICL-NUIM datasets.
- Efficiency: FastORB-SLAM achieved a frame rate of approximately 84 Hz on a 1.10 GHz CPU, making it roughly twice as fast as ORB-SLAM2. Establishing correspondences took only ~12ms, compared to ~16ms for the baseline.
- Accuracy and Robustness: The system maintained high precision across 17 test sequences. Notably, in low-texture environments where unique descriptors are scarce, FastORB-SLAM's intensity-based tracking often outperformed traditional methods.
Conclusion
FastORB-SLAM demonstrates that high-performance visual SLAM is possible on low-power hardware by eliminating redundant descriptor computations. By combining the speed of direct intensity-based tracking with the long-term stability of feature-based mapping, it provides a robust solution for real-time robotics. Future work will focus on integrating IMU data and improving performance in dynamic environments.
Summary updated