Development of software and hardware complex for control and access management based on face recognition algorithms and neural networks

Abstract

The purpose of the study is to develop and test a software and hardware complex for access control and management using face recognition algorithms and neural network technology. A software and hardware complex was used, including turnstiles with swing doors, single-board mini-computers on each turnstile, webcams, sensors, an integrated graphics processor, as well as a software development kit for developers. Siamese neural network “FaceNet” was used for face recognition. When training a neural network, each face image is associated with a feature vector lying on the N-dimensional hyposphere. The study developed and tested photo pre-processing algorithms (to exclude irrelevant images), as well as the biometric PACS algorithm. Approbation of the developed hardware and software system on the campus of the Kemerovo State University (Kemerovo, Kemerovo region – Kuzbass, Russia) showed that the average frame processing speed is high (0.076 seconds per frame), which ensures a fairly fast pass of people. This result is achieved on the basis of the image pre-processing algorithm. During the first test of the software and hardware complex, it was found that 60% of the frames are excluded from the analysis by the neural network. Errors of the first kind (passing people who do not have the right to pass) were observed in 4% of cases. After setting up and calibrating the software and hardware complex, the proportion of such errors decreased to 0.4% of the total number of frames processed by the neural network.