Edge AI: Efficient AI Deployment For Real-Time Capabilities In Autonomous Driving

AI and more specifically Deep Learning have shown remarkable results in various domains integrating multiple facets of our society (Autonomous Driving, manufacturing, medical field...). For the development and deployment of AI-powered products/solutions, many companies rely on the cloud entailing many challenges related to latency, bandwidth, privacy and energy consumption/ CO2 footprint. These challenges raise questions about the reliability of cloud-powered AI products for real-time use-cases such as in autonomous driving schemes (e.g. Pedestrian detection). To answer the above challenges, Edge AI has emerged as an alternative to cloud-centric solutions as it reduces latency, enhances data privacy, minimises bandwidth requirements and aligns more with sustainability needs. The talk will dive into the technical foundations of Edge AI presenting real-life scenarios as well as its importance into enabling efficient AI deployment for real-time capabilities highlighting the contrast to cloud-based AI. Furthermore, the presentation will discuss the challenges that Edge AI faces (limited memory and computational power, heterogeneously-distributed environments) and how to overcome them. A real life use-case will then be presented about the topic of Vulnerable Road User detection discussing an End-to-End pipeline from ethical data generation, efficient training and optimization to deployment on edge. Ultimately, this presentation will advocate for a (partial) shift from the cloud to the edge for efficient and sustainable real-time AI-systems providing the key ingredients to achieve that.Takeaways from the Talk:

​

• Concepts of Edge AI will be presented

• How to optimize AI models for an efficient deployment on edge using open-source frameworks

• The need for more transparency in AI usage w.r.t the environmental impact

• Ethical AI is an important part in Autonomous Driving