MISIS and Citilabs: Algorithm for identifying vehicles in video surveillance systems

Main article: Video analytics (terms, applications, technologies)

2023: Optimizing CCTV algorithms

Scientists of the University of Science and Technology MISIS, together with specialists from the Citylabs company, have improved video surveillance algorithms that determine lubricated and illuminated car numbers. Pre-classification of image quality significantly saves computing resources and improves the accuracy of the entire video surveillance system. The modules are cross-platform, they can be installed on various devices. This development can be successfully used, both on general roads and in some mining facilities. This was announced on September 7, 2023 by the university.

One of the important tasks that arise in the analysis of road traffic situations, including in the conditions of technological roads, is the identification of a specific car by the state registration plate. Often, due to the high speed of the car, the bright headlights, dust content, as well as insufficient capabilities, the cameras of the car are incorrectly recognized. Timely screening of deliberately incorrect number images allows you not to waste computing resources for recognition, and also reduces the likelihood of erroneous recognition.

To determine the degree of illumination of the car number, experts propose to use the analysis of the histogram of brightness. The well-known yolo-v5 neural network is used to detect both vehicles and license plates.

To identify cars and numbers when training neural networks, datacets were formed taking into account the time of day, seasonality and weather. After determining the region of state signs in the image, the selected area from the three-dimensional RGB color space is reduced to a one-dimensional "gray." After counting the histogram, the part of it that will be responsible for "overexposure" is distinguished, thus, 95.7% of the numbers were correctly classified as illuminated. To determine the degree of lubrication, a neural network with architecture was built, which provides a classification accuracy of 96.4% with a minimum processing time of 0.073 ms on a PC, - said Dr. Igor Temkin, Head of the Department of Automated Control Systems (ACS) of NUST MISIS.

A separate task in the course of work on a neural network to determine lubricity was to create a training data set. The conditions under which images are blurry are specific, and it takes a lot of time to select from a huge amount of data those that were suitable for a class of blurry numbers.

In addition to classification into readable and unreadable images, the developed algorithm also provides a quantitative assessment of the degree of lubricity and illumination. This data, in turn, can be used to adjust camera parameters, such as shutter speed and aperture, which will improve the quality of subsequent personnel.

At industrial enterprises, stationary video surveillance systems are quite widely used. The identification of dump trucks based on the analysis of video frames is relevant for controlling the entry, exit and movement of vehicles through the quarries in which non-metallic building materials are mined: crushed stone, sand, gravel. At the same time, rather strict requirements are imposed on the accuracy of number recognition, - said co-author of the development Vladislav Epifanov, graduate student of the Department of ASU of MISIS University.

During the experiments, the proposed approaches have shown their effectiveness on various devices, such as a PC and the Nvidia Jetson Nano microcomputer. The proposed methods are applicable to use both in server solutions and in mobile "boxed" solutions, where the camera and computing device are a single device.