As the healthcare industry evolved, research for providing personalized services became more active. Activity Recognition techniques have been studied from the past as a representative method of recognizing user''s daily life patterns. In addition, due to the development of wearable devices such as Smart Watch, Smart Phone, Activity Recognition technology is being studied using various sensors.
Conventional Activity Recognition system based on Multi-Modal sensor pre-define sensor location and activities, and recognize them using all sensors. Since this method uses all the sensor data, there is a problem that if one sensor is not worn, the activity recognition system does not operate or the accuracy is lowered. Also, this system has a disadvantage in that it is difficult to construct an optimized system that uses different models according to the number of attached sensors, because all the activities to be recognized are defined beforehand and the optimized modeling study is not performed according to the sensor attachment positions.
In this paper, we analyze the most suitable activities according to user ''s sensor location and propose a method to recognize optimal activity through individual activity modeling. In this paper, we propose a method to calculate feature importance by guided random forest algorithm and to calculate the contribution of sensor contributed by each activity using sensor. Also, . In addition, we propose a method to generate an individual activity model using only the appropriate sensor attachment locations and the features for each action. By using only optimized features for each sensor and activity, optimized results for each activity and sensor combination can be derived.
Experimental results show that the recognition accuracy is about 7% higher than that of each existing activity recognition technique.