Computer Vision Aids for the Partially Sighted


The inadequacy of information provided by long canes and the very limited acceptability of guide dogs on the part of partially sighted persons have prompted the development of electronic mobility aids. This motivated the ASMONC project which aims to apply state-of-the-art technology to help the partially sighted to improve their lives. The proposed system uses stereo vision, sonar and infra-red sensors for obstacle detection with GPS and mobile phone for navigation and communication. Obstacle detection and elevation changes detection such as kerbs and steps are two of the most basic requirements of any mobility aid for the partially sighted. This thesis concerns the vision part of the project regarding the detection of obstacles, kerbs and stair-cases.

The ASMONC system is to be carried on the user as a backpack and therefore the stereo cameras will be moving around while the user is walking. The Ground Plane Obstacle Detection (GPOD) algorithm which has been applied successfully in mobile robotics is not appropriate in this case as a one-time fixed ground plane calibration cannot be used to detect small obstacles. It is extended to RANSAC Dynamic Ground Plane Recalibration (DGPR) which recalibrates the ground plane at each frame using random sample consensus. Kalman Filtering is applied to track the ground plane features and obstacle features. Moreover, a technique is developed to match object edges using some heuristic criteria similar to those used in stereo matching. Experiments show that obstacles are detected much better with our dynamic recalibration approach and that objects can be found successfully by the edge matching technique.

Kerbs and stair-cases are useful environmental landmarks that the partially sighted need to be made aware of. Kerbs are detected by identifying clusters of parallel lines using the Hough Transform and from the discontinuity of the ground plane disparity. Stair-cases are detected by looking for groups of concurrent lines, then convex and concave edges are partitioned using intensity variation information. Stair-case pose is estimated by a homography search approach. Using an a priori stair-case model, search criteria and constraints are established to find its vertical rotation and slope. Another method is developed to determine the vanishing line of the stair-case plane, estimating its pose. As a real application, it is crucial to know how reliable the estimation is, therefore, detailed error analysis are carried out for these algorithms which have been applied to both synthetic and real images with promising results.

However, when stair-cases are far away, their edges are not resolvable. Therefore, instead of treating them as individual edges, they are considered as a piece of texture. As a result, texture detection methods using statistics criteria, windowed Fourier Transform and Gabor Filters are employed to recognise distant stair-case regions.