FACIAL RECOGNITION SYSTEM MODEL FOR FIXED VIDEO SURVEILLANCE SYSTEMS

Abstract

To create intelligent video surveillance systems, it is necessary to perform a certain number of tasks related to computer vision. One of the main tasks is the detection of moving objects. This is usually done through pixel segmentation, which allows you to separate pixels of interest from others. These pixels are usually called foreground pixels, while the others are called background pixels. This has led to the development of many approaches by scientists, each of them trying to overcome the challenge in a specific scenario, such as complex background, high illumination, poor quality input image. This process is called background subtraction. The proposed approach will process a collection of images from a video, called frames, and should be fast enough to be applied in real time. Among the mentioned operations there is detection and localization in the case of processing a specific object. If we are dealing with people, the detection will involve localizing the faces in the input frame, which processes the entire frame to find the positions of the faces. In large video surveillance systems, this creates a problem of using resources, especially memory and computing power. The idea is to reduce the number of video sequences to be processed by limiting further computation of the input frame to the region of interest only. This paper proposes an approach that achieves these goals and shows the impact of such operations on resource consumption as well as the accuracy of the results.

Modern video surveillance systems utilize deep learning methods. Such solutions pose many challenges, especially regarding computational resource availability. With the emergence of deep learning approaches, these challenges have become even more complex because deep learning algorithms are based on artificial neural networks and require a lot of specific resources, such as GPUs (Graphics Processing Units). To overcome these challenges, solutions requiring less resource consumption are proposed. This involves adding a motion detection algorithm, which leads to restricting further computations only to the area of interest, i.e., the area containing humans, cars, etc. This article proposes a model for face recognition based on motion detection for stationary surveillance cameras. The study is focused solely on humans and does not cover all types of objects.