The School of Computer Science is pleased to present…
Autonomous mobile robots require reliable obstacle detection to navigate unstructured indoor environments. Standard solutions rely on active sensors such as LiDAR or RGB-D cameras, which impose deployment constraints and increase both cost and power consumption. Passive stereo vision is a low-cost alternative, but it fails in practice due to calibration drift and textureless surfaces found in indoor environments. This thesis introduces a hybrid, uncalibrated stereo framework designed to remain robust under both these challenges. The pipeline performs automated online rectification to correct geometric drift, then fuses sparse texture cues with dense disparity and plane fitting to detect obstacles across both texture and texture-starved regions. A deep visual place recognition module further verifies detections against a reference map, thereby distinguishing static infrastructure from dynamic obstacles and enabling change detection. Experiments on custom datasets demonstrate improved obstacle recovery in texture-starved scenes, where standard stereo approaches often fail.
Internal Reader: Dr. Dan Wu
External Reader: Dr. Mohamed Belalia
Advisor: Dr. Boubakeur Boufama
Chair: Dr. Sherif Saad Ahmed
Registration Link (only MAC students need to pre-register)