Pedestrian Recognition by Using a Kernel-Based Multi-modality Approach

Abstract

Despite many years of research, pedestrian recognition is still a difficult, but very important task. We present a multi-modality approach, that combines features extracted from three type of images: intensity, depth and flow. For the feature extraction phase we use Kernel Descriptors, which are optimised independently on each type of image, and for the learning phase we use Support Vector Machines. Numerical experiments are performed on a benchmark dataset consisting of pedestrian and non-pedestrian (labelled) images captured in outdoor urban environments and indicate that the model built by combining features extracted with Kernel Descriptors from multi-modality images performs better than using single modality images.