Introduction
Vision-Aided Inertial Navigation Systems (VINS) are critical for motion tracking in robotics and augmented reality. However, developers often face a stark trade-off between monocular and stereo configurations. Monocular VINS is computationally efficient but suffers from scale ambiguity, relying on noisy accelerometer data to estimate physical dimensions. Conversely, stereo VINS provides robust scale through simultaneous image pairs but imposes a heavy computational burden and increased latency—factors that can degrade the "motion-to-photon" performance required for immersive applications.
This paper introduces Alternating-Stereo VINS, a novel approach that bridges this gap. By triggering left and right cameras in an alternating sequence and employing an interpolation-based motion model, the system achieves the scale accuracy of stereo vision with the computational efficiency of a monocular pipeline.
Methods
The core innovation lies in the asynchronous triggering and state estimation scheme. Unlike traditional stereo systems that fire both sensors simultaneously, this method alternates between the left and right cameras.
Selective Cloning and Interpolation
To minimize overhead, the system utilizes Selective Cloning. In standard Multi-State Constraint Kalman Filters (MSCKF), the device state is "cloned" at every image timestamp. In this architecture, the system only clones the state for the left camera's timestamps. For observations from the right camera, which occur between left-camera clones, the authors introduce Interpolation-based Jacobians. This linear interpolation model relates right-camera features to the neighboring left-camera clones, assuming relatively smooth motion over the short inter-frame intervals (~60-80ms).
Observability Analysis
A key contribution of the work is a rigorous mathematical proof demonstrating that the absolute scale of the environment remains observable despite the asynchronous nature of the images. The known physical baseline between the cameras (the "stereo constraint") provides sufficient geometric information to resolve scale even when images are not captured at the same instant.
Experimental Evaluation
The system was evaluated using a custom wide-stereo hardware rig (19.3 cm baseline) and the industry-standard EuRoC MAV dataset. Performance was benchmarked on an NVIDIA Jetson TK1 mobile processor to simulate real-world deployment constraints.
Key Results
- Real-Time Efficiency: The alternating-stereo pipeline processed one second of data in approximately 977ms, achieving true real-time performance. In comparison, a full stereo system at the same frame rate required 2609ms, failing to keep pace with the data stream.
- Accuracy: The proposed method consistently outperformed monocular VINS in terms of Root Mean Square Error (RMSE). As shown in the performance analysis, the alternating-stereo configuration (highlighted in the green dashed box) maintains accuracy levels comparable to full stereo systems while operating at a significantly lower computational cost.
Conclusion and Future Work
Alternating-Stereo VINS demonstrates that simultaneity is not a prerequisite for stereo benefits. By mathematically bridging the temporal gap between camera triggers, the system provides a high-performance, low-latency alternative for resource-constrained platforms.
Future research may address the limitations of the linear motion assumption. While effective for typical movements, extremely high-frequency vibrations or jerky motions might require higher-order interpolation models (such as B-splines) to maintain accuracy. Additionally, the requirement for precise hardware-level alternating triggers remains a consideration for integration into off-the-shelf consumer mobile devices.
Summary updated