Real-time Face Analysis System Basing On Dense Crowd

With the rapid development of electronic information technology and computer science,our lives are becoming more and more intelligent,smart retail will be widely used in the near future.When customers enter the shopping mall,the mall's face analysis system can accurately analyze the customer's face attributes and provide customers with appropriate information forwarding so as to guide customers for consumption.Simultaneously,shopping malls using the face analysis system can perform the passenger flow statistics,analyze the customer's product preferences and conduct accurate marketing and so on.Therefore,the analysis of face attributes has a very wide range of application prospects.This paper focus on the urgent need for this practical application,designs and realizes a real-time face analysis system for dense crowds.It not only designs a compact,fast and accurate face recognition algorithm,but also combines traditional machine learning and deep learning methods to implement the algorithm on the development board.The main works of this paper can be concluded as follows:(1)Algorithm design for face analysis system: creatively combines traditional machine learning algorithms and deep learning algorithms to make full use of the speed advantage of traditional algorithms and the accuracy advantage of deep learning algorithms;for the face detection part,there is a detection algorithm based on multiblock LBP face features and a cascaded AdaBoost classifier,and a detection algorithm performs CUDA parallel acceleration based on GPU;for the face attributes analysis part,a multi-task micro network implementation algorithm based on deep learning is proposed.(2)Implementation of real-time face analysis system: In this study,combined with the embedded system NVIDIA JETSON TK1,a small CNN network was designed to achieve face attribute analysis.In the case of ensuring accuracy,the speed was significantly improved.(3)The system's experimental case analysis: according to long-term analysis of a large number of faces in a real scene,a series of conclusions have been drawn,which verifies that the system does have a very high practical value.The system can be used to analyze audiences in real time,and to deliver precisely matched advertising information accordingly.It can also be used in media effect measurement and market research analysis.This achievement will not only promote a new field of face analysis,but also solving a general problem.