On the performance of an IoT edge-based computer vision solution to monitor agglomerations

Abstract

The Internet of Things (IoT) has attracted more attention currently due to its concept of connecting physical objects on the internet and making them intelligent in such a way that they can process the information of environmental, human and other objects. It enables to make cities smart and helps to solve problems in traffic and energy systems. The home equipment can be automated and hospitals and clinics can share the data of patient quickly and securely. Agriculture can improve its production due to better soil care and optimization of natural resources. Public security, which is an important foundation of society, can also benefit from the use of IoT and create mechanisms that make citizens comply with laws. And it is in this context which the computer vision can be used conjunction with IoT, as this combination allows to monitor agglomerations in public places through the detection of people and assisting authorities in controlling the crowd. However, cloud architecture in this IoT scenario has drawbacks in terms of security, latency and bandwidth. As a solution, the processing of image should be shifted to the edge of the network, near where the surveillance cameras are installed, and to transmit only the result of the processing. This architecture is known as Edge Architecture and it proves to be an efficient option to process the sensitive information close to the origin. This dissertation proposes the deployment of a new edge scheme which uses computer vision to detect people in agglomerations. Based on this, an enhancement in a state-of-the-art algorithm is proposed to perform the monitoring in equipment of limited resources. To evaluate the performance, a comparative analysis of detection is carried out using four videos from different environments and in the presence of external interference. The results obtained demonstrate gains in terms of accuracy and computational performance on video analysis in comparison to the state-of-the-art solutions available in the literature, in addition to increasing the privacy of sensitive information and reducing the occupation of network resources through local processing