articles

Information about store traffic is of great value to businesses. It allows you to evaluate the effectiveness of marketing campaigns and optimize staff work schedules. Moreover, data on the number of visitors can be indirectly used to analyze the competitive environment. Despite the existence of various technological approaches to solving the problem of counting visitors, each of them has a number of significant disadvantages. The purpose of the research is to develop a software system for counting visitors based on the application of machine vision technologies to a video stream. To do this, it was proposed to split the counting task into two subtasks: detection and tracking of visitors’ movements in the frame, each of which was solved using convolutional neural networks. Training and validation of neural networks were carried out on data collected in real conditions exclusively from cameras of the system customer. Together with the advanced counting algorithm, the system became capable of: a) excluding from the count employees of the retail chain wearing a corporate uniform; b) correctly handles complex and unpredictable trajectories of visitors in a video surveillance scene; c) without compromising the calculation accuracy, correctly handle video stream decoding errors, which result in dropped frames. Testing of the quality of the system’s operation was carried out on 504 test videos, in which a total of 739 visitors entered and left the outlet. When processing each frame, the final calculation error was 3%. And in the course of a number of experiments, it was found that when processing only every 4 frames (the load on the system in this case was reduced by 4 times), the calculation error increased by only 1%.