Detection of pedestrians for far-infrared automotive night vision systems using learning-based method and head validation

To improve active automotive safety and guarantee the protection of pedestrians at night time, a fast pedestrian detection approach based on a monocular far-infrared camera for driver assistance systems is presented. According to the distribution of gray-level intensity of pedestrian samples, an adaptive local dual threshold segmentation algorithm is executed first to extract candidate regions. Then a novel entropy weighted histograms of oriented gradients (EWHOG) descriptor is proposed for effective pedestrian description in a far-infrared spectrum. For reducing the within-class variance of pedestrians located at different distances, a three-branch classifier combining EWHOG features and a support vector machine (SVM) is introduced for robust pedestrian recognition. To speed up the recognition phase, the resulting support vectors (SVs) are optimized to reduce the number of SVs used for decision-making. A further validation phase is then introduced to suppress the false alarms according to the intensity difference between pedestrians’ heads and their adjacent regions. Comparisons between our approach and conventional approaches are presented, and experimental results show that the presented pedestrian detection framework is very promising.