Real Environments Image Labelling Based on Reconfigurable Architectures

In this paper a vision system of images perceived in real environments based on reconfigurable architecture is presented. The system provides surface labelling of the input images of unstructured scenes, irrespective of the environmental lighting or scale conditions and the capture frequencies. The nucleus of the system is based on querying SOMs constructed with supervised training by means of descriptors extracted from images of different surfaces perceived for successive values of the optical parameters (lighting, scale). To improve the labelling process the system estimates the environmental optical parameters and then SOMs are reconfigured. A segmented architecture is proposed for the central module of the labelling process to improve the performance of image sequences. A prototype implemented on FPGAs applied as a guidance aid for vehicles is provided.