High-Speed Stereo Visual SLAM for Low-Powered Computing Devices

Abstract: We present an accurate and GPU-accelerated Stereo Visual SLAM design called Jetson-SLAM. It exhibits frame-processing rates above 60FPS on NVIDIA's low-powered 10W Jetson-NX embedded computer and above 200FPS on desktop-grade 200W GPUs, even in stereo configuration and in the multiscale setting. Our contributions are threefold: (i) a Bounded Rectification technique to prevent tagging many non-corner points as a corner in FAST detection, improving SLAM accuracy. (ii) A novel Pyramidal Culling and Aggregation (PyCA) technique that yields robust features while suppressing redundant ones at high speeds by harnessing a GPU device. PyCA uses our new Multi-Location Per Thread culling strategy (MLPT) and Thread-Efficient Warp-Allocation (TEWA) scheme for GPU to enable Jetson-SLAM achieving high accuracy and speed on embedded devices. (iii) Jetson-SLAM library achieves resource efficiency by having a data-sharing mechanism. Our experiments on three challenging datasets: KITTI, EuRoC, and KAIST-VIO, and two highly accurate SLAM backends: Full-BA and ICE-BA show that Jetson-SLAM is the fastest available accurate and GPU-accelerated SLAM system (Fig. 1).

• The paper introduces Jetson-SLAM, a GPU-accelerated stereo visual SLAM system that achieves over 60 FPS with minimal trajectory error on low-powered devices.
• Bounded Rectification and PyCA techniques optimize feature detection and selection, reducing misclassification and computational overhead.
• The novel Frontend–Middle-end–Backend design minimizes CPU-GPU data transfers, enabling efficient real-time mapping on embedded platforms.

High-Speed Stereo Visual SLAM for Low-Powered Computing Devices

The research presented in the paper introduces Jetson-SLAM, a GPU-accelerated Stereo Visual Simultaneous Localization and Mapping (SLAM) system optimized for performance on low-powered computing devices, like NVIDIA's Jetson-NX. The study addresses the challenges faced by low-powered devices in processing-intensive SLAM tasks, especially in the stereo configuration that demands high computational resources for both front-end and back-end processes.

Key Contributions

The authors identify three primary contributions to the Jetson-SLAM framework:

1. Bounded Rectification Technique: This enhancement to the FAST feature detection method seeks to prevent the misclassification of non-corner points as corners, thereby improving SLAM's accuracy. By refining feature detection, the method reduces statistical outlier impact due to noise or minor blobs that are incorrectly classified as corners.
2. Pyramidal Culling and Aggregation (PyCA): PyCA consists of Multi-Location Per-Thread (MLPT) culling and Thread-Efficient Warp-Allocation (TEWA) schemes that leverage GPU capabilities for fast and efficient feature selection. MLPT optimizes the culling of features from image pyramids by varying scale and resolution, while TEWA ensures high GPU throughput by reducing resource wastage even on devices with limited cores.
3. Frontend--Middle-end--Backend Design: A novel middle-end SLAM component is introduced to handle stereo-matching, feature-matching, and feature-tracking with optimized data sharing. This reduces the overhead associated with CPU-GPU data transfers by utilizing a synchronized shared memory mechanism, thus minimizing computational resource usage and improving runtime performance.

Results and Evaluation

Jetson-SLAM showcases its capabilities through extensive evaluation across multiple datasets like KITTI, EuRoC, and KAIST-VIO, demonstrating both high processing speeds and substantial accuracy improvements. Experimental findings highlight that Jetson-SLAM achieves significant frame rates exceeding 60 FPS on Jetson-NX with reduced computational load while maintaining minimal trajectory error across challenging sequences. Particular attention is drawn to PyCA's ability to drastically reduce feature counts without sacrificing accuracy, resulting in effective reduction of runtime for SLAM pipelines.

The validation against benchmarks shows a marked improvement over existing VO/VIO/SLAM systems, particularly in scenarios with fast motion or low light. Jetson-SLAM achieves these improvements while utilizing resources efficiently on embedded GPU platforms.

Implications and Future Directions

Jetson-SLAM represents a pivotal advancement in SLAM system design for low-powered devices, expanding the potential for real-time applications in autonomous systems where resources are constrained. The techniques introduced optimize computational tasks by aligning GPU parallel processing with efficient algorithms, potentially influencing future low-power real-time embedded architecture developments.

Looking ahead, Jetson-SLAM's integration into wider robotic systems presents opportunities for enhancing autonomous unmanned aerial vehicles (UAVs), robotics controllers, and edge computing use cases where traditional SLAM systems are impractical due to resource limitations. Further advancements may focus on adapting these strategies for broader types of SLAM systems beyond stereo configuration, enhancing scalability and functionality to encompass more diverse sensing inputs and environmental interactions.