• Speaker : Prof. Sanmin Kim
• Date : May 28, 2026
• Affiliation : Kookmin University
• Category : Invited Seminar

Vision-based 3D Perception for Autonomous Driving

Abstract

While autonomous driving relies heavily on various sensors, camera-based vision systems offer crucial semantic information and cost-effectiveness. This seminar provides a deep dive into 3D perception technologies using camera images, ranging from foundational concepts to the latest trends in End-to-End Autonomous Driving (E2E AD). Prof. Kim discusses key methodologies including 3D Object Detection and dense 3D Occupancy Prediction voxel grids for understanding complex driving environments, alongside emerging frameworks like Vision-Language-Action models and Planning-aware Perception.

Presentation Overview

This presentation covers the following key topics:

• 3D Object Detection (3DOD) evolution from monocular and stereo images to multi-view 3D detection to resolve spatial-temporal ambiguity.
• 3D Occupancy Prediction using dense voxel grids and self-supervised pre-training to handle irregularly shaped obstacles and minimize labeling costs.
• The transition to End-to-End Autonomous Driving (E2E AD) paradigms via World Models and Planning-aware Perception to optimize results.