Robust pedestrian detection for driver assistance systems using machine learning

Abstract

Vision-based pedestrian detection is a challenge task for a variety of applications such as driving assistance systems, especially in case of insufficient illumination. Effective fusion of complementary information acquired by multispectral images (visible and infrared) allows robust pedestrian detection under various lighting conditions (e.g., day and nighttime). In this paper, we propose a multispectral pedestrian detection approach that combines visible and infrared images. Firstly, an Otsu thresholding is applied to infrared images to detect hot spots most likely representing a pedestrian, after applying some morphological operations to enhance the original image and compensate for clothing-based distortions. The significant regions of interest obtained in the infrared image are mapped into corresponding visible image. Secondly, multispectral aggregated channel features are used with a thermal discrete cosine transform, as descriptor combined with a support vector machine (SVM) classifier. Our approach is evaluated on the KAIST multispectral dataset to prove its efficiency.