Computational Method for Recognizing Situations and Objects in the Frames of a Continuous Video Stream Using Deep Neural
- PDF / 3,389,643 Bytes
- 16 Pages / 612 x 792 pts (letter) Page_size
- 83 Downloads / 147 Views
ERN RECOGNITION AND IMAGE PROCESSING
Computational Method for Recognizing Situations and Objects in the Frames of a Continuous Video Stream Using Deep Neural Networks for Access Control Systems O. S. Amosova,*, S. G. Amosovaa, S. V. Zhiganovb, Yu. S. Ivanovb, and F. F. Pashchenkoa a
Institute of Control Sciences, Russian Academy of Sciences, Moscow, Russia State University, Komsomolsk-na-Amure, Russia *e-mail: [email protected]
b Komsomolsk-na-Amure
Received January 13, 2019; revised May 1, 2020; accepted May 25, 2020
Abstract—An effective (performance- and accuracy-wise) computational method for pattern recognition in a continuous video stream using deep neural networks for access control systems is proposed. The class of recognition problems solved by the method using a sequence of video stream frames is identified: the vehicle itself and the characters on its license plate (LP), faces of people, and abnormal situations. In contrast to the known solutions, a classification with a subsequent reinforcement based on multiple frames of a video stream and with an algorithm for the automatic annotation of images is used. Neural network architectures with independent recurrent layers for classifying video fragments adapted for the problems, a dual network for face recognition, and a deep neural network for vehicle character recognition are proposed. New databases for neural network training are created. A schematic diagram of an intelligent access control system for ensuring the security of an enterprise, a distinctive feature of which is the use of a multirotor unmanned aerial vehicle with a computing unit, is proposed. Field experiments are carried out, and the accuracy and performance of the computational method in solving each problem are assessed. Software modules in the Python language for solving tasks of the intelligent access control system are developed. DOI: 10.1134/S1064230720050020
INTRODUCTION Computer vision systems (CVS) are currently increasingly introduced in various domains. Patterns of physical and technical objects, as well as situations involving them, can be captured in video stream frames. The set of properties is different for each object or situation. Examples of the properties for technical objects, such as vehicles, are the type, number, color, etc., and for physical objects, such as people, the gender, age, etc. The properties of objects can be assessed by individual frames, even by one of them. Situations differ significantly from objects because the former are characterized by duration and the relationships between dynamic objects. Therefore, it is necessary to use a sequence of frames of a continuous video stream to assess the properties of situations. The main task of computer vision algorithms is to search for patterns in the image, identify their key features, which characterize the properties of objects and situations, and recognize them for further decision making or control. Recently, the use of deep neural networks (NNs) in CVS recognition problems has become one of the most nota
Data Loading...
