Locating and monitoring people in indoor controlled environments : a robust depth-based approach

Abstract

The majority of the population within developed countries is living longer than ever before and evidence is signalling that old adults prefer to stay in their homes and communities. Having said that, the reality is posing huge challenges for ageing in place, hence there is an increasing demand for solutions to this problem, most of which revolve around technology. Projects like NATIFLife (Network of Assistive Technology for an Independent and Functional Life) aim at developing a framework of assistive technology for the domestic environment, to support the elderly and people with disabilities to lead an independent life. This work focused on the design, implementation and installation of a localisation and monitoring system which locates and tracks people within an indoor environment. An ambient intelligence approach was adopted whereby usability is always on top of the agenda. Hence, none of the technology should be obtrusive or something that involves a steep learning curve. There should be no manual for participating in ambient intelligent environments - it should just work as you walk in. The prioritisation of usability led to focus this work on only vision-based technology as it remains the most important modality for providing rich information related to Human Behaviour Understanding in an unobtrusive manner. Moreover, due to the nature of the application and environment in which the system seeks to operate, privacy was a major concern hence for this reason, the operation of the system was restricted to using depth imaging only. The proposed system consists of various building blocks amongst which the main are; calibration, detection and tracking modules. A custom trained YOLOv3 neural network was chosen as the technique for conducting human detection in depth imagery, that operates in real-time on low-end hardware within a network of depth cameras and computing machines. The system achieves an average of 71% detection accuracy, 59.6% tracking accuracy (single subject) within the set limitations.