Vision-based Resource Constrained Event Detection for Medical Smart Homes
Mahdy Nabaee · MacSphere (McMaster University) · 2010
As the number of elderly persons as well as their fraction of the total population continues to rise, especially in the developed countries, providing an appropriate living environment for them using smart home technology is rapidly gaining attention. Two important tasks of a smart home technology are monitoring the daily activities and the vital signs of the elderly to improve their quality of life and to monitor existing or the onset of health abnormalities. In this thesis, we focus on the monitoring of taking medicine by the elderly person using vision sensors (low-cost cameras). This task is important since it helps both the person and the doctor in the treatment of illnesses of elderly persons. The allocated resources of communication bandwidth between the sensor nodes and the computational power, used for this task, affect the implementation cost. Therefore, it is desired to develop an effective scheme which efficiently allocates bandwidth and computational resources to achieve a high reliability (detection performance) at low cost. In this thesis, we have proposed two different approaches to solve this detection and monitoring problem. As the input data are video frames, captured by cameras from the same scene, the frames have inter-view redundancy. Taking advantage of this inter-view redundancy, we proposed a video coding classification scheme based on separate encoding and joint decoding, and have obtained significant compression improvement compared to existing techniques. In the second approach, we studied different parts of the detection and monitoring system to find an efficient design for distribution of different event detection parts between the nodes and the central processing unit so that the allocated resources are reduced. In this scheme, the useful information of the frames are extracted in the form of their main features such that decision making based on these features is the same as decision making based on the raw frames. As a result, we could propose a new scheme which requires significantly less bandwidth and computational resources while achieving the same detection performance.