An Obstacle Categorization System for Visually Impaired People

In this paper, we introduce a new framework for obstacle localization and classification. The proposed method is designed to improve cognition of visually impaired people (VI) facilitating the autonomous navigation in outdoor environments. In the context of computer vision applications, the following contributions are proposed and validated: (i) a new method of selecting a reduced and relevant set of interest points, (ii) a novel descriptor denoted Adaptive Histogram of Oriented Gradients (A-HOG) dedicated to arbitrary categories of objects, (iii) an image re-ranking method at Vector of Locally Aggregated Descriptor (VLAD) / Bag of Visual Word (BoVW) descriptor level based on graph spanning structures and neighborhood relations. Finally, we demonstrate the performance of the proposed framework (in terms of classification accuracy and computational time) on a challenging video dataset captured with the help of real VI users. The entire framework is completely integrated on an Android smartphone device, while all methods were specifically designed and tuned under the constraint of achieving real-time processing capabilities.