Fusing Semantic Segmentation and Object Detection for Visual SLAM in Dynamic Scenes
summary
A Visual SLAM system fuses a semantic segmentation branch and an object detection branch with a self‑adaptive switch and a geometric static point recovery (SPR) step to keep true static features inside detection boxes. Evaluated on TUM RGB‑D, it improves pose estimation versus DS‑SLAM by at least 13% on high‑dynamic sequences, runs about 10.3 fps, and avoids tracking failure modes seen when removing all box features or using slower methods like DynaSLAM.
AI-written summary · about
chapters (4)
related talks
26mFrom Panoramic Vision to Lightweight 3D Worlds for Semantic Visual Tracking
13mEnhancing Object Detection & Avoidance Systems for Robots in Challenging Environments Using AI
17mSwarmMap: Scaling Up Real-time Collaborative Visual SLAM at the Edge
31mDeep learning for self-driving vehicle technology
49mUsing Deep Learning for Perception in Autonomous Systems: A Perspective from Control Theory
10mREGROUP: A Robot-Centric Group Detection and Tracking System
30mUAV-based Person Detection and Geolocation for Search and Rescue
26mAI for Autonomous mobile robots