Optimizing Change Detection for Autonomous Mobility: Insights from Mobileye’s REM™ System
Overview
This post, authored by Mobileye’s engineering team, details our innovative approaches and architectural decisions in enhancing road change detection via the Road Experience Management (REM™) system. We also express gratitude to our collaborators at AWS.
The Importance of High-Definition Maps
Our high-definition maps are essential for precise vehicle navigation and localization, providing a foundation for the autonomous driving capabilities of today and tomorrow.
Change Detection in REM™
Understanding and adapting to changes in road structure is critical for autonomous vehicles. This section outlines the complexities involved and the mechanisms we employ for effective change detection.
Cost Efficiency as a Priority
In designing our deployment solution, we prioritized cost efficiency over other metrics, utilizing Amazon EC2 Spot Instances for a scalable and fault-tolerant approach.
Architectural Innovations
We share our experiences in optimizing the Change Detection pipeline by transitioning to CPU-based deep learning inference, leveraging Triton Inference Server, and adopting AWS Graviton instances for enhanced performance.
Results Achieved
Through our optimizations, we have significantly improved the efficiency and throughput of our Change Detection system, enhancing the overall user experience.
Future Outlook
The journey toward optimizing our systems is ongoing, and we anticipate further advancements to bolster performance and cost efficiency in the realm of autonomous mobility.
Driving Change: How Mobileye’s Road Experience Management™ Harnesses AI for Safer Roads
By Chaim Rand, Pini Reisman, and Eliyah Weinberg at Mobileye
In our rapidly evolving transportation landscape, ensuring autonomous vehicles (AVs) can navigate the complexities of real-world conditions is paramount. At Mobileye, we’re committed to leading the charge toward smarter, safer mobility through advanced technologies like the Road Experience Management™ (REM™) system—a critical framework for creating high-definition (HD) maps essential for precise vehicle localization, real-time navigation, and adaptability to changing road conditions.
Understanding Road Experience Management™
REM™ allows Mobileye to gather and maintain extensively crowdsourced HD maps that serve as the backbone for autonomous driving. These maps enable AVs to pinpoint their location with unprecedented accuracy while continually adapting to environmental changes, from new road constructions to alterations in lane markings.
The evolution of effective mapping hinges on continuous data processing from millions of vehicles equipped with our technology, making it imperative to develop efficient, scalable solutions capable of managing such computationally intensive tasks.
Focus on Change Detection
One key area of our REM™ system is Change Detection—an automated process that identifies alterations in road structures. As the ancient philosopher Heraclitus famously noted, “the only constant in life is change.” This principle is especially relevant to road networks. Whether a new lane is added during construction or a detour is instituted, inaccurate mapping can pose significant challenges for AVs, emphasizing the importance of rapid and reliable map updates.
Change Detection operates on multiple road segments globally, utilizing a proprietary algorithm to analyze data collected from Mobileye-equipped vehicles. Our deep learning model, CDNet, plays a pivotal role in this process, generating insights that inform map updates as necessary.
Key Decisions and Trade-offs in Deployment
1. CPU vs. GPU: Striking the Balance
Initially, we explored leveraging GPU instances for the CDNet model’s execution due to their superior performance. While GPUs do offer significant speed advantages—processing up to 54.8 samples per second compared to 5.85 on CPUs—we found that utilizing CPU instances yielded better overall cost efficiency.
The high demand for GPU instances often leads to increased costs and lower availability—critical factors for our continuous and large-scale change detection needs. By shifting to CPU instances, we eliminated unnecessary overhead associated with resource management, realizing significant gains in a system designed for efficiency over raw speed.
2. Centralizing Inference with Triton Inference Server
To counter memory limitations and processing bottlenecks, we adopted Triton Inference Server for centralized model inference. This optimization reduced our per-task memory requirement from 8.5 GB to 2.5 GB and halved average task runtimes, allowing for the execution of up to 32 tasks per instance—a clear boost in efficiency.
3. Diversifying with AWS Graviton Instances
The introduction of AWS Graviton instances marked a transformative step in bolstering our change detection capabilities. These ARM-based processors provide outstanding price-performance ratios and are tailor-made for machine learning workloads. By integrating Graviton instances into our pool, we not only enhanced throughput efficiency but also expanded our instance availability, vital for meeting peak task demands.
The Results: Enhanced Throughput and User Experience
The impacts of our architectural refinements are significant. For instance, by employing AWS Graviton instances, we observed improvements in processing capacity that doubled our overall efficiency. Our deployment accounted for user experience enhancements by significantly amplifying the Change Detection system’s throughput, thus fostering confidence and reliability in AV navigation.
Additionally, the seamless integration of Graviton’s architecture into our existing system exemplifies how modern frameworks can adapt rapidly to evolving technological landscapes.
Closing Thoughts
The journey toward optimizing change detection is ongoing. At Mobileye, we recognize that machine learning frameworks are continuously evolving, offering opportunities to enhance performance and flexibility. This endeavor not only fosters innovation for safer vehicles but also reinforces our commitment to leading the future of intelligent mobility.
Acknowledgments
We extend our gratitude to Sunita Nadampalli and Guy Almog from AWS for their invaluable contributions to this endeavor.
In the ever-changing world of road ecosystems, our work at Mobileye serves as a reminder that leveraging AI and cloud technologies can drive remarkable advancements, ultimately paving the way for safer journeys for all. Stay tuned for more insights and developments as we continue to refine and innovate our strategies in this exciting realm.