Majority of the elderly, especially in developed countries, living on their own and they are considered as an "at-risk" group. They appear to be associated with higher risks of accidental falls that are reported as the most common cause of injury for the elderly. The instant treatment for injuries of fallen people is very critical. The degree of injury is proportional to the delay time in receiving medical treatments. Hence, falls should be detected as soon as possible since accidental falls seem to be unavoidable and hard to be predicted. Timely responses help fall people not worsen the injuries.
In the health care industry, there is a tremendous demand for supportive products and technologies to improve the safety at home. One typical example is Personal Emergency Response System (PERS) composing of a small radio transmitter, a console connecting to the user's telephone, and an emergency response center that monitors such types of calls. In emergency cases, for example, fall events, users can press the "HELP" button to contact with the emergency response centers to receive immediate assistance. However, the conventional PERS tends to be impractical in assisting the elderly living alone since "HELP" button must be carried all the time. The elderly may easily forget to carry it due to the dementia or the deterioration of cognitive ability. Moreover, the impact of shock after accidental falls may force fallen people to experience unconscious states of mind as well as physical pain. Pressing "HELP" button to call for immediate assistance seems to be impractical solution. Thus, an "Intelligent" PERS that is capable of providing automatic sensing of emergencies is in demand for the health care of the elderly.
There are different technologies of fall detection but can be categorized into three mainstreams.
The first one uses, for example, wearable accelerators to measure the accelerators of the human body. But it is intrusive in the subject's daily life. The second one uses, for instance, floor vibration detectors, requiring a complicated setup and still being in their infancy, albeit being promising solutions in the future. Meanwhile, the last one is the most promising and practical solution that is not only to detect fall events accurately but also preserve the privacy and facilitate the freedom of the elderly. Here we only consider the vision-based solutions.
In this project, we develop "Multiview 3D spatial feature - based approaches to fall detection."
The First Method:
We realized that heights and occupied areas of people in usual states, i.e, standing and sitting and in suspicious states, i.e, lying are significantly different. We propose to approximate the person of interest by 3D cuboids to extract its heights and bottom areas as occupied areas. Human states such as standing, sitting and lying are classified according to the features of heights and occupied areas by SVM. The 3D cuboids are simply reconstructed by using 2D bounding boxes extracted from two cameras whose fields of view are relatively orthogonal. These 2D bounding boxes are normalized by appropriate Local Empirical Templates, which are defined as templates of standing people in the vicinity of the people. The normalization cancels the perspective effect and makes the features invariant across the viewing window. Given a sequence of human states, a fall event is detected by analyzing human state transition. We test the proposed method on 'Multiple Camera Fall Dataset' and obtain competitive results with state-of-the-art methods, tested on the same dataset, but iwith lower computational cost.
Publications:
The second Method:
By using heights and occupied areas, we can distinguish lying from standing and sitting states. However, the information of where the person lies either on the ground or on a sofa is unknown. Consequently, the previous method is only able to detect a state change from standing to lying as a fall. Sit-to-stand transfer falling type in which people change states from sitting to lying is not considered. Therefore, we propose using contact area between people and the ground, so-called Human-Ground Contact Areas (HGCA). People make a little contact with the ground during usual activity but often lie completely on the ground after suffering from accidental falls. By using HGCA, we can know whether people lie on the ground or on a sofa. We propose a low-cost scheme of estimating HGCA from multiple views. Foreground of people are projected from one view to another by using homography of the ground between views. The homography of the ground suggests that only foreground of body parts which are contact with the ground, for instance, the feet, are consistently projected across views. We measure the overlap regions between the foregrounds of the people and the projected foregrounds from the other views as HGCA.
We also propose a 3D human state simulation in which a virtual camera placed on the surface of a hemisphere to capture 3D human models in various poses from a variety of angles. Hundreds of images are generated from the simulation as training data. HGCA are extracted from the training data to find a threshold to separate lying states from standing and sitting states. Given a sequence of human states, a fall can be detected by temporal analysis of human state transition. The performance of the proposed approach on 'Multiple Camera Fall Dataset' are favorably comparable with state of the art methods, tested on the same dataset, but with lower computational cost.
Publications: