A Real-time Human Activity Recognition Approach with Generalization Performance

The current human activity recognition (HAR) methods need training data from users. The data collection causes discomfort to the users and most of the studies ignore the real-time performance of classification. This paper presents a real-time human activity recognition approach with strong generalization performance. It uses existing dataset to avoid long-term data collection of subjects, so that the machine can be quickly applied to each specific individual. Also, it takes advantage of both combined accuracy and limited feature selection proposed by this paper to implement feature-selection-based transfer learning which improves HAR in both real-time and generalization performance. In view of the recognition time and accuracy, the depth neural network is selected, changeable structure of which is more suitable for feature selection. This approach utilizes four inertial measurement units placed on the outside of human thighs and shanks. A total of seven activities are taken into account that includes level-walking, upstairs, downstairs, uphill, downhill, standing and sitting. The experiments are performed on six healthy male subjects in free-living settings to evaluate the efficacy of the algorithm. This approach achieved a notable activity recognition accuracy of 98.89%, and reported a fast average activity classification time of 28.6 ms.