Abstract

Built a monocular-camera obstacle-avoidance pipeline for the DOGZILLA quadruped, estimating scene depth with ORB-SLAM and segmenting obstacles from U/V-disparity maps — enabling real-time autonomous navigation in unstructured environments without a dedicated depth sensor (LiDAR or stereo). Led the four-person RoboStride team through Korea University’s Creative Challenger Program; the project won the President’s Award (1st Place) and was later presented at the KSPE 2025 Spring Conference.

Method

• Estimated scene depth from a single RGB feed with monocular ORB-SLAM, extracting sparse 3D point clouds for obstacle localization
• Built U-disparity and V-disparity maps from the estimated depth to segment obstacles from the ground plane by disparity peaks
• Fed detections into the robot’s navigation controller to trigger real-time avoidance maneuvers

Platform

DOGZILLA quadruped · monocular RGB camera

Skills

C++ Python ROS2 ORB-SLAM OpenCV