| Abstract [eng] |
Observing social distance is essential in maintaining personal health and safety, as it is one of the essential preventive measures against airborne viruses and diseases. Therefore, research is constantly being carried out to create technological solutions to ensure accurate social distance detection in public spaces. The main existing security methods are based on single-camera images analysed by convolutional neural networks. These models identify the people in the frame and transform the camera view into a "top-down" perspective, allowing the space of each detected person and the distance to the surroundings to be determined. The main goal of this study is to apply a multi-view pedestrian detection algorithm to a social distance monitoring system. Furthermore, the impact of the spatial placement of cameras on tracking accuracy and other important detection indicators is also investigated. In addition, existing social distance monitoring solutions and their algorithms are reviewed. The EarlyBird algorithm chosen for the research uses images from several cameras, extracting their essential features, performing perspective transformation, and combining the obtained results to determine the most likely positions of individuals. Improvements made during the study allow a straightforward selection of available cameras for evaluating the quality of social distance monitoring. Finally, potential improvements to the initial stage of the algorithm are explored using more advanced convolutional neural network models for person detection. This research demonstrates that a multi-view pedestrian detection system significantly improves accuracy in detecting social distancing violations. Moreover, utilising a more complex backbone for feature map extraction did not substantially enhance accuracy. However, using the less complex TinyNet-E model resulted in faster training and inference times, with only a marginal reduction in accuracy. |
