Embedded Face Analysis for Smart Videosurveillance

Abstract

In this paper, we describe our methodology for designing a smart Videosurveillance system for face analysis. The system aims at increasing the security by gathering demographic statistics in highly crowded areas such as train stations, airports and shopping malls. Based on Convolutional Neural Networks (CNNs), the system architecture relies on the reconfigurable hardware to accelerate part of the computation and reduce the power consumption compared to general-purpose processors and GPUs. To achieve easy programmability, the platform makes use of the OmpSs programming model, which provides parallelization and acceleration by using simple directives to be added to the sequential code. The rsource-intensive tasks are offloaded to the reconfigurable hardware in order to achieve the desired performance levels. Our evaluation shows that we can detect more than 600 faces per frame, while keeping the power consumption at about 8W. The tests were performed by using the AXIOM hardware/software platform.