SVM-based pedestrian recognition on Near-InfraRed images

This paper describes the algorithms we developed for a new automotive night vision system for pedestrian detection based on Near Infrared (NIR) illuminators and sensors. The system applies in the night domain the SVM technique, which has already been successfully implemented in day-light applications. In this project we have developed optimizations in order to meet accuracy and time performance requirement for in-vehicle deployments. In particular, we present a novel pre-SVM processing technique, which performs pixel-level and multi-resolution analysis in order to discard portions of the frame that are not likely to contain pedestrians. This procedure allows exploiting the SVM as a very accurate classifier focused on the most critical cases.