Multiple Pedestrian Tracking Framework using Deep Learning-based Multiscale Image Analysis for Stationary-camera Video Surveillance

Abstract

A novel single static-camera multiple pedestrian detection and tracking system, which could be succesfully used by the Smart City technologies, is introduced in this article. The moving person detection process is performed by applying a combination of advanced computer vision and machine learning solutions, such as Gaussian Mixture Models (GMM), Histogram of Oriented Gradients (HOG), Support Vector Machines (SVM) and Aggregate Channel Features (ACF), to each frame of the color video sequence. An instance matching-based tracking technique that uses a deep learning-based multiscale analysis of the subimages of the detected pedestrians is then proposed. Its scale-space is created by applying the numerical approximation algorithm of a well-posed nonlinear anisotropic diffusion-based model that is introduced here. The results of the pedestrian detection and tracking experiments are described in the end.