Introduction
Multi-user Augmented Reality (AR) requires devices to quickly and accurately localize themselves in a shared physical space. This is a computationally intensive task typically handled by Simultaneous Localization and Mapping (SLAM) algorithms. For a seamless experience, multiple users must interact with virtual holograms anchored to the same real-world coordinates, requiring each device to estimate its pose (tracking) and contribute to a shared 3D representation of the environment (mapping).
Running full SLAM on mobile devices is often too slow and power-hungry, while offloading to a distant cloud introduces prohibitive latency. This paper evaluates various networked architectures—from peer-to-peer to edge-cloud-centric—to determine how best to distribute these computations to meet the strict latency and spatial accuracy requirements of diverse AR applications.
Architectural Frameworks for Multi-User SLAM
The authors identify three primary networked architectures for distributing SLAM tasks:
- Edge-Cloud Centric (e.g., SLAM-Share): Both tracking and mapping are offloaded to the Edge Cloud (EC). The EC leverages GPU parallelization for feature matching and a shared memory buffer to merge maps from multiple users instantly. Clients stream video to the EC and receive lightweight pose coordinates in return.
- Edge-Cloud Assisted: Devices perform tracking locally but offload mapping to the EC. The EC merges data and sends updated global maps back to the clients. However, large map updates can be sensitive to network congestion.
- Peer-to-Peer (P2P): Devices handle all SLAM tasks locally and share map updates directly with one another, bypassing centralized infrastructure.
Performance and Use Case Analysis
A critical finding is that the Edge-Cloud Centric model is significantly more resilient to network lag than the Edge-Cloud Assisted model. In assisted setups, network delays starve the client of heavy map updates, causing local tracking to degrade. In contrast, the centric model only transmits tiny pose updates. Short transient delays in the centric model can be effectively smoothed using the device's local Inertial Measurement Unit (IMU).
Experimental results showed that under a 300 ms round-trip delay, the Edge-Cloud Centric architecture maintained a stable Average Tracking Error (ATE), while the Edge-Cloud Assisted model saw significant error spikes. Even with delays up to one second, the centric model's absolute trajectory error increased by less than 1 cm.
The "best" architecture is ultimately dependent on the specific requirements of the application, as summarized in the classification matrix below:
- High-Precision Needs: Surgical training requires 10 ms latency and mm-level accuracy, making it strictly reliant on Edge-Cloud Centric models.
- Moderate Requirements: Autonomous vehicles can utilize Edge-Cloud Assisted architectures, as they require sub-second latency but can tolerate meter-level accuracy.
- Casual AR: Applications like Pokemon Go can function effectively using P2P architectures due to more relaxed latency and accuracy constraints.
Conclusion and Future Directions
While the Edge-Cloud Centric approach offers superior performance for high-stakes AR, it requires reliable, high-bandwidth communication. Future research must address the need for hybrid fallback mechanisms. In scenarios where edge infrastructure is unavailable or network links fail, systems should gracefully transition to local or P2P computations to ensure continuous safety and usability.
Summary updated