Development and preliminary analysis of sensor signal database of continuous daily living activity over the long term

A new corpus of daily living activities using wearable sensors and a living activity recognition method based on sensor signals are presented. The corpus consists of both indoor and outdoor living activities measured by a small camera and a smartphone over 72 continuous hours. We collected sound and image data from the camera and motion signals from the smart phone. We then analyzed the sensor signals and performed experiments on living activity recognition using a Gaussian mixture model based on the sensor signals. Experimental results showed that combining acoustic and motion features with weighted likelihood can improve recognition accuracy compared to utilizing acoustic features only. This demonstrates the effectiveness of integrating acoustic and motion features to recognize daily living activities.