Class binarization with self-adaptive algorithm to improve human activity recognition

Flourishing research in Human Activity Recognition (HAR) is essential in improving the quality of an individual’s health. Low cost and privacy interest, sensing technology becomes an imperative topic in activity monitoring applications. Nevertheless, the presence of high interclass similarity from similar activities mainly involving stairs activities yields to degrade the recognition accuracy. These kind of activities highly sparsely distributed in the input space which is problematic to be distinguish using traditional classifier model. Even though deep learning becomes a recent imperative topic, model complexity is considered as a foremost drawback and impractical to be conducted. Furthermore, although better recognition of stairs activities is accomplished, recognition of stationary activities is less reported due to less sensitivity of lesser waveform. Somehow, it might occur some of extracted features are insignificant to describe the activity. Even if a ranking method is widely utilized in solving numerous of dimension reduction problems such as in bioinformatics and high spectral images, most of works are disregarding the boundary to discard the irrelevant features. In order to improve recognition of high interclass similarity activities, One-Versus- All (OVA) binarization strategy is introduced by transforming original multi-class classification problems into a series of two-class classification problems. However, the learning complexity of classification is increased due to the expansion number of learning model. Therefore, feature selection using Relief-f with self-adaptive Differential Evolution (rsaDE) algorithm is proposed to select the most significant features. To enhance the selection of most highly ranking features, irrelevant features are ‘pruned’ based on determined boundary threshold. In order to estimate the quality of ‘pruned’ features, self-adaptive DE algorithm is proposed. Two parameters (population size and generation numbers) are adaptively adopted from number of remaining ranking features. Also, self-adaptive scaling factor and crossover probability control parameters are introduced to diminish time of finding an optimal parameter to produce the best population. In order to investigate the correlation between features and class, generated feature subsets are rearranged according to its mutual information. In such circumstances, frequency domain features are proposed due to their less susceptible to signal quality variations and beneficial to recognize stationary activity. These features are combined with statistical features to improve the ability of classifier model in distinguishing between locomotion, stationary and complex activities. Two publicly activity datasets are used; Wireless Sensor Data Mining (WISDM) and Physical Activity Monitoring for Aging People (PAMAP2). WISDM consists of six different types physical activity, while PAMAP2 covers eighteen activities comprising various simple and complex activities. In comparison, WISDM utilizes an accelerometer sensor embedded in Android smartphone. Meanwhile, PAMAP2 utilizes an accelerometer sensor equipped with three Inertial Measurement Unit (IMU) devices attached to three different placements. Performance of the proposed method is compared with several benchmark works. Experimental results have significantly promised an improvement of activity recognition level, mainly involving very similar activities.

Text FSKTM 2018 68 - IR.pdf Download (740kB)

Actions (login required)

View Item