Wearable sensor technology is evolving in parallel with the demand for human activity monitoring applications. According to World Health Organization (WHO), the percentage of health problems occurring in the world population, such as diabetes, heart problem, and high blood pressure rapidly increases from year-to-year. Hence, regular exercise, at least twice a week, is encouraged for everyone, especially for adults and the elderly. An accelerometer sensor is preferable, due to privacy concerns and the low cost of installation. It is embedded within smartphones to monitor the amount of physical activity performed. One of the limitations of the various classifications is to deal with the large dimension of the feature space. Practically speaking, a large amount of memory space is demanded along with high processor performance to process a large number of features. Hence, the dimension of the features is required to be minimized by selecting the most relevant feature before it is classified. In order to tackle this issue, the hybrid feature selection using Relief-f and differential evolution is proposed. The public domain activity dataset from Physical Activity for Ageing People (PAMAP2) is used in the experimentation to identify the quality of the proposed method. Our experimental results show outstanding performance to recognize different types of physical activities with a minimum number of features. Subsequently, our findings indicate that the wrist is the best sensor placement to recognize the different types of human activity. The performance of our work also been compared with several state-of-the-art of features for selection algorithms.

Accelerometer; Differential evolution (D); Evolutionary algorithm (EA); PSO; Genetic algorithm (GA); Particle swarm optimization (PSO); Relief-f; Tabu search algorithm

Human Activity Recognition (HAR) application has recently gained attention in the intelligent environment field. In such states, monitoring human activity might be extremely important to reduce the fraction of unhealthy conditions. According to the 2016 report from the World Health Organization (WHO), the percentage of diabetes patients has increased incrementally in the world population (WHO, 2016).

In looking at this matter, insufficient physical activity is one of the issues. Regular exercise can be thought as one of the simplest solutions, by spending time in at least twice a week engaged in some physical exercises. Also, with the advancement of sensing technology, the use of inertial sensors offers a possible solution. Inertial sensors, such as an accelerometer and gyroscope, provide opportunities to undergo the HAR application process and these sensors have also been equipped in various smartphone models. Hence, everyone can track and monitor their daily exercise without relying on any other additional devices. These micro-machine electromechanical systems (MEMs) sensors are able to record the signals in three-dimensional spaces, where the x-axis (left-and-right movement), the y-axis (up-and-down movement) and the z-axis (upward/backward movement) are monitored (Acharjee et al., 2015). The signal is recorded by quantifying the sense of  vibration through the device when movement is triggered. Even so, the choice of the sensor placement also effects to the classification performance (Avci et al., 2010). So, the position of the sensor placement needs to be clearly identified which to ensure it is able to recognize different kinds of actions, particularly in recognizing complex activities. These complex activities could be considered as the activity which consists of a sequence of actions arising from several different parts of the human body.

Also, to deal with the abundance number of features is another challenge. Practically speaking, the training model complexity and the processing time is strongly related to the numbers of features to be processed (Catal et al., 2015). In such states, feature selection was ‘pruned’ by removing the less significant features before each feature is classified. The features that do not contribute enough information to be described within the particular class are removed from the feature space.

In this article, several contributions were carried out. A hybrid feature selection method using Relief-f feature ranking with a well-known evolutionary algorithm, known as differential evolution (DE) was proposed to select the most relevant features. Secondly, we also proposed an adaptive parameter mechanism without relying on the exhaustive process to find the optimum parameter values. Lastly, our proposed feature selection method also performed an outstanding degree of accuracy which was better than several well-known feature selection algorithms, such as particle swarm optimization (PSO), evolutionary algorithm (EA), genetic algorithm (GA) and the Tabu search algorithm. This paper is organized as follows: Section II explains the background work; Section III describes the materials and methods; Sections IV discusses the proposed feature selection.; Section V presents the results and discussion; Section VI presents the conclusions.