#8 How a self-driving car learns how do drive with Saquib Sarfraz (Mercedes Benz-Daimler)

2022-01-31 · Source: AI LITERACY - A Podcast about Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Novice, extended

Summary

Saquib Sarfraz, Senior Scientist Computer Vision at Mercedes-Benz Daimler, discusses the evolution and future of autonomous vehicles and computer vision. He highlights that while deep learning has accelerated performance, core computer vision research for tasks like object recognition has existed since the 1990s. The current surge in autonomous driving capabilities is largely due to increased computational power and the vast availability of real-world driving data collected from vehicles equipped with sensors like cameras and LiDAR. Sarfraz explains that autonomous systems mimic human observation and prediction using neural networks, which learn to map inputs (like images) to outputs (like object identification) through layers of artificial neurons. He details how a simple filter can detect boundaries and how neural networks learn these filters at multiple levels to recognize complex objects and predict intent, such as a pedestrian crossing the street. Despite rapid technological advancements, the full rollout of Level 5 autonomous vehicles faces significant non-technical hurdles, including regulatory approvals, insurance frameworks, and infrastructure adaptation, making a widespread deployment unlikely within the next 15-20 years.

Key takeaway

For Computer Vision Engineers developing autonomous driving systems, recognize that while the technology for Level 4 and some Level 5 capabilities exists, widespread adoption is constrained by regulatory, insurance, and infrastructure challenges. Focus your efforts on robust, verifiable systems that can progressively integrate into existing frameworks, understanding that full autonomy will be an evolutionary process requiring significant societal and infrastructural shifts over decades, not just technological breakthroughs.

Key insights

Autonomous vehicles leverage neural networks and vast real-world data to mimic human observation and predictive intelligence.

Principles

• Intelligence is observation and prediction.
• Neural networks learn mathematical input-output mappings.
• Evolutionary change requires infrastructure adaptation.

Method

Neural networks learn to detect features (e.g., boundaries, shapes) by adjusting internal "filters" (weights) across multiple layers, enabling hierarchical reasoning for object recognition and intent prediction from sensor data.

In practice

• Utilize real-world sensor data for robust model training.
• Implement layered neural networks for complex perception tasks.
• Consider non-technical factors for AV deployment timelines.

Topics

• Autonomous Vehicles
• Computer Vision
• Neural Networks
• LiDAR Sensors
• SAE Autonomy Levels

Best for: Computer Vision Engineer, Research Scientist, AI Student, Software Engineer, AI Product Manager

Related on AIssential

• Xiaomi EV’s Next Self-Driving AI Explained — Autonomous driving models can achieve superior performance by shifting from linguistic reasoning to a latent spatial-temporal mathematical representation.
• How Self-Driving Technology Completely Changed — Self-driving technology rapidly shifted from rules-based to neural network architectures, making prior investments largely obsolete.
• Scania: AI-Driven Accuracy Underpins Autonomous Systems — AI-driven autonomous freight technology offers precision and safety to transform supply chain logistics and mining operations.
• Self-driving cars struggle to see at night or in fog – but imitating the human brain can make them safe — Biologically inspired divisive normalisation significantly enhances AI vision system robustness in adverse conditions.
• Artificial Intelligence for Modeling and Simulation of Mixed Automated and Human Traffic — AI methods are crucial for accurately simulating complex mixed human and autonomous traffic behaviors.
• The System Behind Self-Driving: Waymo’s Dmitri Dolgov — Full autonomy demands a comprehensive AI ecosystem, integrating advanced sensing, simulation, and specialized models for robust, scalable deployment.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI LITERACY - A Podcast about Artificial Intelligence.