A driver assistance system based on multilayer iconic classifiers: Model and assessment on adverse conditions

Recent work demonstrates that iconic classifiers are good candidates for the development of effective driver assistance systems, exploiting on-board micro cameras and embedded architectures. Following this line of research, in this paper the combined use of multilayer classifiers and iconic data reduction, based on Sanger neural networks, is investigated. It is shown that by this affordable approach it is possible to capture the essential information of the images, making worthless much more structured and time-consuming feature-based techniques. In particular, the applicability of a simplified learning stage, based on a small dictionary of poses, is considered; this peculiarity makes the system almost independent from the actual user. A detailed model of a simple driver assistance system, based on iconic classifiers, is presented and a comparative assessment, focused on the specific task of monitoring the car driver, is performed on adverse driving conditions. Three well known classification techniques are applied, demonstrating that the iconic approach, though can be certainly improved, is characterized by robustness, accuracy and real-time response; these features prove this technology to be an ideal tool for embedded automotive applications.