The COVID-19 pandemic highlighted the importance of contact tracing to limit the spread of infections. The problem with tracing systems that require the use of apps and other approaches, which burden users, is that they often go unused outside of a pandemic. This master's thesis introduces a system for identifying close contacts, offering a non-invasive solution for contact tracing within buildings. The system comprises two parts: the camera surveillance system allows for real-time monitoring of social distancing between individuals and can be easily integrated into existing camera systems. The close-contact detection system enables us to identify all the close contacts of a specific person, but only when necessary. This conserves resources and maintains the privacy of visitors. For the purposes of detection and identification of individuals, an efficient model named Eff-SeqNet was developed, which can be used on graphical processing units available to end users. With this, we demonstrated that by making a thoughtful choice in model architecture, a good balance between detection accuracy and speed can be achieved. In addition to the new efficient model, we introduced a person search pipeline utilizing information about position and connections between cameras, which improves the robustness as well as the efficiency of our system.
|
