Kyungseo Park, Yong Lin, Jyothi Vinjumur
Using wireless sensor networks, we would like to detect abnormal human activities for the elderly who lives alone. Using existing techniques, such as longest common subsequences, top-k ranking, Euclidean distance similarity search, and decision boundary, we would like to classify normal and abnormal activity based on low level sensor data. The data additionally include temporal information so that we can detect abnormal activity caused by temporal aspects. After the project is completed, we can have online system that monitors human activity and warns whenever abnormal activities are detected.
Figure 1: A testbed one bedroom apartment with sunSpots sensors are installed at
4-digit hexadecimal numbers indicate unique sensor IDs. We used light and accelerometer sensors to detect human activity.
Figure 2: Preliminary result to classify normal and abnormal behavior.
X axis represents time and Y axis represents normalized score. This shows score that determines normal and abnormal activity based on a 14 days training set. We tested 5,10, and 15 events that are included in a series, which makes an episode. By using decision boundary technique, we need to determine a good point to classify normal and abnormal activities.