The School of Computer Science is pleased to present…
Vision-Based Path Monitoring for Human Walkers
MSc Thesis Defense by: Robert Odoh
Navigation within indoor environments remains a significant challenge for individuals with visual impairments, particularly when obstacles appear unexpectedly or when the walking path becomes difficult to discern. Traditional electronic travel aids often rely on fixed infrastructure, specialized sensors, or coarse location information, limiting their practicality and scalability. This thesis presents a real-time, vision-based obstacle-awareness system designed to support safe and independent indoor mobility without requiring additional environmental instrumentation.
The proposed system integrates a lightweight YOLO object detector with an automatically calibrated homography-based ground-plane model. This combination enables the system to determine not only what objects are present but where they lie relative to the user’s walking path. A no-click floor-estimation module derives the geometry of the ground plane from the camera feed, allowing detected objects to be projected into metric space and evaluated based on their position and distance. A path-aware filtering pipeline then distinguishes between objects that lie safely outside the walking corridor and those that present an immediate collision risk. In addition, a flat-object suppression mechanism prevents thin, floor-aligned items (such as papers or magazine covers) from being mistakenly treated as obstacles.
The system operates in real time on consumer hardware and requires only a single forward-facing camera. Experimental testing across varied indoor scenes demonstrates its ability to identify true obstacles, suppress irrelevant floor patterns, and provide consistent spatial awareness without the need for markers, depth sensors, or user-initiated calibration. This work contributes an infrastructure-free, computationally efficient solution that enhances indoor mobility and supports safer, more confident navigation for visually impaired users.