Learning hierarchical spatial semantics for visual orientation devices

Complexity of understanding a visual scene is the single biggest challenge in creating intelligent devices for visually impaired people. The requirement of real time operation makes it inevitable to design algorithms that obey the computing and memory limits of available hardware. We present a hierarchical scene understanding system implemented on a vision system chip. It is restricted to extract specific information for predefined categories of visual scenes, but it is general enough to be able to learn quickly and autonomously. Patches having potential discriminative information are extracted using a hierarchical peeling method. Object groups are created based on proximity and size of the patches. Objects are classified using different classifiers and the votes are combined using a mixture of experts network. Experimental validation has been carried out on authentic image flows recorded by blind subjects.