Proposed FPGA hardware architecture for high frame rate (>100 fps) face detection using feature cascade classifiers

Abstract

Face detection is the first and most crucial step in any face recognition systems with applications to face tracking, recognition and automatic surveillance. AdaBoost based face detection training methods have been proposed for producing a rapidly fast detection implementation compared to other methods using the integral image for evaluating a series of weak classifiers very fast. Current software version implementations can achieve about 15-25 frames per second (fps) with a tunable compromise between detection accuracy and speed. In this paper, a novel hardware architecture design on FPGA based on AdaBoost face training and detection algorithm for detecting faces in high resolution images at high frame rates (>100 fps) is proposed. The proposed architecture can evaluate each sub-window in a single clock cycle, which is the fastest possible speed land is limited by the H/W clock running speed. Thus in this proposed approach, detection speed is independent from the number of weak classifiers implemented. The proposed architecture is verified on Xilinx Virtex-II Pro FPGA platform where experimental results show that the speed and memory outperform previous approaches in literature, achieving 143 fps for image size of 640 by 480 pixels using a single scan window. Parallelizing the scan window can lead to double/triple this speed and is dependent on the gate capacity of the FPGA.