Abstract
This work presents a comprehensive study of the effectiveness of computer vision methods for the task of detecting faces in images and in video streams. The main attention is paid to the comparative analysis of two algorithms of the dlib library: the classical method based on histograms of oriented gradients (HOG) and the modern method based on convolutional neural networks (CNN). The problem of high computational complexity of neural network methods is considered and a solution is proposed through the use of NVIDIA CUDA parallel computing technology. A software algorithm for benchmarking is developed, which allows for automatic evaluation of frame processing speed (FPS), inference time and detection stability on different hardware. The experiment proved that the use of a graphics processor (GPU) allows achieving multiple acceleration (speedup) of image processing when using CNN, providing the ability to work in real time with high accuracy. The results of the study allow us to determine the optimal hardware and software configurations for building video surveillance, access control, and biometric identification systems. The conclusions obtained can be used in the design of high-load video analytics systems.

This publication has 13 references indexed in Scilit: