Video Processing Algorithms for Detection of Pedestrians

In this paper a video processing procedure for automatic detection of pedestrians is presented. It is planned to use it as a part of the automotive night vision system. Generally, such systems are either passive (i.e. those based on thermal vision) or active (i.e. equipped with illuminators and near infrared cameras). Passive systems provide a large range of detection, while their active counterparts, operating in a somehow smaller range, offer more readable images for car drivers. However, all images produced with both kinds of these systems are quite specific and special image processing procedures are needed for them. For this purpose the authors used modified and adapted algorithms, such as dual-threshold locally adaptive classification, connected component labeling, histogram of oriented gradients, and the support vector machine with a radial basic function kernel or with a linear kernel. Tests performed on the real night vision recordings show very high efficiency of the proposed solution with accuracy equal to 99.2% for the linear kernel and even to 99.36% for the radial basic function kernel.