From Panoramic Vision to Lightweight 3D Worlds for Semantic Visual Tracking
summary
Claims a GPS-free urban localization pipeline that uses a 360 panoramic camera plus lightweight, textureless 3D worlds generated from public sparse point clouds to localize and track mobile robots. Uses semantic segmentation to keep only building pixels, then aligns real and rendered views with frequency-domain Gaussian-mixture optimization and an adaptive elliptical Gaussian mixtures regulator to handle large rotations and occlusions. Demonstrated in Unity and on a Renault Zoe car with better accuracy than SLAM/optical-flow baselines, published at ICRA 2026 and used in an Amazon collaboration.
AI-written summary · about
chapters (11)
- 0:00Introduction and background
- 1:05Why GPS fails robots
- 4:37Virtual environments for localization
- 6:21Panoramic vision vs virtual cameras
- 8:09Visual odometry and alignment
- 10:58Gaussian mixtures for tracking
- 13:01Fixing convergence errors
- 15:48Closed-loop trajectory tracking
- 17:53Demonstration and results
- 19:58Industry impact and future work
- 21:00Q&A session
related talks
41mComputer Vision for Remote AR
26mAI for Autonomous mobile robots
10mMapping sidewalks and street furniture with AI and computer vision
09mFusing Semantic Segmentation and Object Detection for Visual SLAM in Dynamic Scenes
17mData Challenges with 3D Computer Vision #bbuzz
16m"Putting a New Spin on
08mUnified Perception: Models for Autonomous Driving
31mDeep learning for self-driving vehicle technology