NIR Image based Pedestrian Detection in Night Vision with Cascade Classification and Validation

Abstract

Pedestrian detection is one of the vital issues in advanced driving assistance applications. It is even more important in nighttime. This paper presents a robust algorithm for a nighttime pedestrian detection system. A NIR (Near Infrared) camera is used in this system to take images of a night scene. As there are large intra class variations in the pedestrian poses, a tree structured classifier is proposed here to handle the problem by training it with different subset of images and different sizes. This paper discusses about combination of Haar-Cascade and HOG-SVM (Histogram of Oriented Gradients-Support Vector Machine) for classification and validation. Haar-Cascade is trained such that to classify the full body of humans which eliminates most of the non-pedestrian regions. For refining the pedestrians after detection, a part based SVM classifier with HOG features is used. Upper and lower body part HOG features of the pedestrians are used for part based validation of detected bounding boxes. A full body validation scheme is also implemented using HOG-SVM when any one of the part based validation does not validate that particular part. Combination of the different types of complementary features yields better results. Experiments on test images determines that the proposed pedestrian detection system has a high detection rate and low false alarm rate since it works on part based validation process