Abstract
In recent years, sensor-based human activity recognition (HAR) has gained tremendous attention around the world with a range of applications. Instead of using body sensor network-based recognition systems which are intrusive and increase equipment cost, we focus on the development of efficient HAR approach based on a single triaxial accelerometer. In order to improve the recognition accuracy of the system, a novel recognition approach based on kernel discriminant analysis (KDA) and quantum-behaved particle swarm optimization-based kernel extreme learning machine (QPSO-KELM) is proposed. KDA is utilized to extract more meaningful features and enhance the discrimination between different activities. To verify the effectiveness of KDA, three kinds of features including original features, linear discriminant analysis (LDA) features and KDA features are extracted and compared for activity recognition. In addition, QPSO-KELM is compared with two existing classification methods: support vector machine (SVM) and extreme learning machine (ELM), which are commonly utilized in activity recognition. Meanwhile, two comparative optimization methods for KELM are also discussed in the experiment. The experimental results demonstrate the superiority of the proposed approach.
Highlights
human activity recognition (HAR) has become an active research area with a wide range of applications
In order to obtain the optimal parameters of the KELM, we optimize C in Eq (19) and the parameters of the kernel function by QPSO
The original features and linear discriminant analysis (LDA) features are utilized for comparison with kernel discriminant analysis (KDA) features
Summary
HAR has become an active research area with a wide range of applications. HAR can automatically recognize activities of daily living by using cameras or inertial sensors and machine learning algorithms. With the development of deep learning techniques, some vision-based HAR methods utilizing convolutional neural networks for feature extraction and activity recognition have been reported [8], [9]. One drawback of these vision-based approaches is that the performance will be greatly affected by external conditions, especially when there is a bad illumination condition. The proposed activity recognition approach is detailed, where the workflow, experimental acquisition equipment, data acquisition and preprocessing, feature extraction method, KDA, and the QPSO-KELM classifier are presented in detail.
Published Version (Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have