165 / 2019-06-18 17:20:37

FPGA Based Real-Time Multi-Face Detection System With Convolution Neural Network

Face detection, CNN, FPGA

The AdaBoost-based real-time face detections have been widely used in current video surveillance. However, the AdaBoost-based face detection has poor performances in detecting multi-face with different scales, multiple poses, and occlusion in complex lighting environment. Recent research shows that the convolutional neural network (CNN) can improve its accuracy. In this work, a FPGA based real-time multi-face detection system for crowded area surveillance application using CNN is presented. A hardware friendly fully quantization strategy is proposed and the result is tested on WIDER FACE dataset. With acceptable loss of accuracy, the FPGA based system can achieve a frame rate of 37 FPS at 512×288 resolution with only 65 ms processing delay.