Detection of pedestrians at night time using learning-based method and head validation

To improve automotive active safety and guarantee the safety of pedestrians at night time, a fast pedestrian detection method based on monocular far-infrared camera for driver assistance systems is proposed. 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. The presented pedestrian detector uses histograms of oriented gradients (HOG) as features and support vector machine (SVM) as classifier. In order to speed up the classification phase, the resulting support vectors (SVs) obtained by SVM is optimized to reduce the number of SVs used for decision. A further validation phase is then introduced to filter the false alarms according to the distribution of gray-level intensity of pedestrians' heads. Experimental results show that the proposed method performs as fast as 34 frames per second on average and guarantees a real-time pedestrian detection; the whole system produces a detection rate of 84.83% at the cost of less than 4% false alarm rate on suburban scenes while produces a detection rate of about 81% at the cost of lower than 10% false alarm rate on urban scenes.