1. Introduction
Human activity recognition (HAR), as a new research area in the field of pattern recognition, has become a topic of focus for many scholars. During the past decade, HAR, especially the activities of daily living (ADL) such as walking, sitting, lying, jumping, and so forth, has attracted much attention from researchers worldwide. Various HAR systems have been proposed by researchers as a medium to obtain additional information about people’s activities. By analyzing the information from patients’ activities, doctors have been able to diagnose some chronic diseases [
1] as well as develop rehabilitation plans for Parkinson’s patients [
2]. Thus, HAR can provide the elderly with better-quality healthcare. Moreover, HAR is also important for applications including human–computer interaction, surveillance, keeping track of athletic activities [
3], and so on.
There have been many solutions to HAR. These can be roughly divided into three aspects: video-based, environment interactive sensor-based, and wearable sensor-based solutions [
4]. For video-based solutions, thermal cameras and depth cameras such as Microsoft Kinect and Intel SR300 have been utilized in activity recognition and experienced great breakthroughs. For example, Xia et al. found that it was feasible to use the Kinect for respiratory motion tracking [
5]. Qin et al. presented a novel method for real-time markerless hand gesture recognition from depth images [
6]. In [
7], thermal and depth cameras have been utilized for multimodal detection of breathing patterns. A review article on both handcrafted and deep-learning-based action representations for vision-based HAR was presented in [
8]. It compared these two approaches and presented the well-known public datasets available for experimentation. In general, video-based solutions perform well if, in a well-controlled environment, it is especially suitable for security (e.g, intrusion detection) and some tracking applications. One problem of video-based solutions is that people’s privacy may be violated if the cameras are installed in some places, such as bathrooms and bedrooms, which limits their application area. Moreover, the performance of video-based solutions may be not robust and reliable if there is clutter or variable lighting in the environment. Last, but not least, video-based solutions are relatively expensive [
9,
10]. Environment interactive sensor-based solutions would not violate the subject’s privacy and are suitable for recognizing daily living activities in rooms. However, this approach is generally costly due to the numerous sensors deployed in appropriate places and is often limited to indoor scenarios [
11,
12]. This fact hinders such a real-time HAR system from being scalable.
With the current development of the microelectronics, miniature and flexible sensors such as accelerometers, gyroscopes, proximity sensors, humidity sensors, and so forth, bring convenience to the users [
13,
14,
15,
16]. Compared with video-based solutions, the wearable sensor-based approach has the advantage of being light and compact, which allows it to collect people’s motion information all the time and anywhere. This kind of approach is also suitable for both indoor and outdoor environments. Therefore, the wearable sensor-based approach can be a good candidate for human activity recognition.
Most HAR studies have utilized sensors from multiple body positions to collect human activity information for recognition [
17,
18,
19]. These systems achieve good recognition results with indistinguishable activities. However, they are not suitable for long-term applications because multiple sensors can cause inconvenience to users. Comparatively, a small number of studies have utilized a single sensor attached to a subject’s body part, such as the waist, chest, or ankle, to collect activity information [
20,
21,
22,
23]. This approach is suitable for long-term activity monitoring and achieves good recognition results for some basic activities such as lying, walking, and running. However, it is not reliable when dealing with complex and similar activity types such as going downstairs and upstairs. Moreover, it may work even worse when processing different data recordings of similar activity caused by individual differences [
24,
25]. Despite significant research efforts to find out the most effective feature selection and feature transformation methods for single-sensor-based HAR, improvements in the robustness and generalization of activity recognition problems with large data variations are still very limited.
As a pattern recognition problem, two aspects could make HAR challenging. First, different subject-related features such as gender, age, weight, and height make HAR a complex problem. For example, the adult and the elder do not have the same kind of data when they are walking or running. Second, the variety of styles with which people perform a certain activity under different external environments is another challenge [
26,
27]. These problems require the recognition system to not only have good recognition accuracy but to also have good generalization ability. Among the methods that can improve the generalization and robustness performance of recognition systems, combinations of multiple classifiers have been demonstrated to be very effective [
28,
29,
30]. However, due to the redundancy between the base classifiers, there is no guarantee that there will be a good complementary relationship between all base classifiers, and some basic classifiers may not contribute to improving system performance. Selecting base classifiers with excellent performance and complementarity can further improve the performance of the recognition system [
31,
32,
33]. This approach is similar to optimizing feature sets to reduce feature dimensions and obtain a more robust feature set. However, as we know, there are very few studies that have applied this classifier approach for constructing a recognition system in HAR.
To address these issues, in this paper, we present a two-layer diversity-enhanced multiclassifier recognition method. Since accelerometers are commonly used and have been proven to be effective for human activity recognition [
29,
34], we only used accelerometer data in this study. Three kinds of features, including time-domain features, autoregressive (AR) coefficients and frequency-domain features, were extracted from a sliding window of data. During the experiments, we first investigated the performance analysis of the effect of training set diversity enhancement on activities that are easily misrecognized using the whole set of base classifiers. Then, we explored the structure and performance of each base classifier in the whole set of base classifiers. Last, we validated the effectiveness of our proposed classifier selection method in activity recognition through a number of comparative experiments.
The main contributions of this paper are summarized as follows:
(1) A novel multiclassifier recognition framework that considers training set and base classifier diversity was proposed to enhance the generalization performance of the HAR system and improve the system’s adaptability to different individuals.
(2) A kernel Fisher discriminant analysis (KFDA) was performed to process the extracted features to enhance the discrimination between different activities, and bootstrap resampling was utilized to increase the generalization performance of classifiers by creating diverse training sets.
(3) A classifier selection approach was applied to the field of activity recognition for the first time, and a novel classifier selection method was proposed to optimize the performance of a multiclassifier recognition system.
(4) We demonstrated that the proposed method is reliable and accurate for activity recognition by collecting sensor data from subjects with a great diversity of subject-related features and comparing the performance of the proposed method to some traditional ensemble methods.
The paper is organized as follows: In
Section 2, related works that focus on feature selection and multiclassifier recognition systems in activity recognition are introduced. In
Section 3, we present details of the proposed activity recognition approach. Following that,
Section 4 introduces the experiment and results. Finally, we draw conclusions in
Section 5.
2. Related Work
The fundamental problem of recognition algorithm design is how to improve the generalization ability and robustness of the recognition system. The performance of the recognition system can be improved by integrating multiple learning individuals that meet certain conditions. Many researchers currently utilize multiclassifier schemes to improve the accuracy of HAR. Catal et al. [
29] established an activity recognition model based on a machine learning classifier. It combines the J48, multilayer perceptron (MLP), and logistic classifiers using the mean method, and the average recognition rate of the model is 97%. Lee et al. [
35] proposed a hybrid expert model based on a smart device to recognize human activity that had a recognition accuracy of up to 92.56%. Yuan et al. [
36] utilized the output of multiple speed learning machines to perform simple mean algorithm fusion processing, and the final model output recognition accuracy was 6% higher than that of a single speed learning machine. Cao et al. [
37] optimized the deployment of multisensors by pruning the multiple ensemble classifier. Through the proposed method, the number and type of multisensor can be appropriately decided. Bayat et al. [
38] built an ensemble-learning-based HAR model which contained three classifiers: MLP, SVM (Support Vector Machine) and LogitBoost for in-hand phone position. This model achieved 91.15% accuracy when recognizing six activities. The experiments also showed that the average of the probabilities was better than majority voting as a fusion method for the proposed model. An ensemble model was built in [
39] using AdaBoost in combination with the decision tree algorithm C4.5 and other base classifiers. The study found that the AdaBoost–C4.5 ensemble model achieved a higher overall accuracy level of 94.04%. Although the above studies have improved the recognition accuracy of the classification system, the individual differences between the classifiers have not been considered. The classifiers participating in the multiclassifier should not only satisfy the accuracy but also must have certain diversities.
Some advanced works have concentrated on feature studies in HAR. Ronao et al. [
40] applied data mining technology to mobile phone sensor-based activity recognition and a deep convolutional neural network was utilized as an automatic feature extractor and classifier. However, this method requires a relatively large data processing capability for the hardware device. Some studies utilized feature transformation methods to reduce the feature dimension while also enhancing the distinguishing ability of feature vectors. For example, [
22,
41] introduced linear discriminant analysis (LDA) to enhance the discrimination between different activities and make features more robust to be useful for fast activity recognition. In order to reduce the influence of the sensor’s varying locations and orientations on the recognition performance, principal component analysis (PCA) was employed in [
42] to realize location-adaptive activity recognition. Wang et al. [
43] proposed a hybrid feature selection method to reduce feature dimensions. This method combined the traditional feature selection method filter and wrapper. The experimental results showed that the method fully balances the relationship between recognition efficiency and accuracy. Motivated by the success of the weightlessness feature, Tao et al. [
44] proposed a new two-directional feature for bidirectional long short-term memory (BLSTM) for incremental learning in human activity recognition. Experiments on the naturalistic mobile-device-based human activity dataset suggested that it is superior to other methods. Forster et al. [
45] proposed a feature extraction method based on genetic programming to obtain a feature set that is robust to sensor position. Experiments on a fitness activity dataset showed that the method achieved an accuracy of 73.4% in contrast to 70.1% when using one selected standard feature. Wang et al. [
46] presented a game-theory-based feature selection method to select distinguished features and reduce computational cost. The experiments showed that the proposed method performed better when compared with ReliefF and mRMR.
Table 1 summarizes these activity recognition studies using wearable sensors on multiclassifier schemes and features.
5. Conclusions
The inertial signals obtained by the same kind of activity under different conditions (e.g., different environments and individual differences) exhibit different characteristics. The absence of a fixed and standardized activity poses a challenge to activity recognition. In order to improve the recognition accuracy and increase the generalization performance of recognition algorithms, this paper proposed a novel activity recognition approach using a single triaxial accelerometer.
There were three critical components in our proposed approach. First, after extracting three different kinds of features from the acceleration sensor, the feature set was mapped to the new subspace by using KFDA technology to enhance the degree of discrimination of feature vectors under different activities. Second, for activity recognition, we proposed a multiclassifier system which contained ELM as the base classifier trained by the bootstrap technique. Third, the base classifiers trained by the bootstrap technique were ranked and selected based on their performance and diversity before combination. Comparative experiments with PCA- and FDA-based features showed that KFDA-based features can improve the classification accuracy effectively. In addition, it can be concluded from the experiments that the base classifiers have different structures and performances in activity recognition problems when they are trained by the bootstrap technique. Based on this, the proposed classifier selection method was utilized to optimize the classifier ensemble and showed a superior advantage compared with combining all base classifiers and the random selection method in the experiments. Apart from these comparative experiments with random selection, the proposed method also showed better performance with traditional ensemble methods, including Bagging and Adaboost.
As future work, we plan to use the EMG acquisition function of the sensor used in this paper and multisignal fusion technology to build a feature set and construct a multiclassifier recognition system with different kinds of classification algorithms. Additionally, we will attempt to use compressed sensing and deep learning methods, such as the deep belief network and the convolutional neural network, to construct two-layer diversity-enhanced activity recognition modules and engage in testing our proposed method by using datasets obtained from more body positions.