| Summary: | Following the recent advances in technology and the growing use of mobile devices such as smartphones, several solutions may be developed to improve the quality of life of users in the context of Ambient Assisted Living (AAL). Mobile devices have different available sensors, e.g., accelerometer, gyroscope, magnetometer, microphone and Global Positioning System (GPS) receiver, which allow the acquisition of physical and physiological parameters for the recognition of different Activities of Daily Living (ADL) and the environments in which they are performed. The definition of ADL includes a well-known set of tasks, which include basic selfcare tasks, based on the types of skills that people usually learn in early childhood, including feeding, bathing, dressing, grooming, walking, running, jumping, climbing stairs, sleeping, watching TV, working, listening to music, cooking, eating and others. On the context of AAL, some individuals (henceforth called user or users) need particular assistance, either because the user has some sort of impairment, or because the user is old, or simply because users need/want to monitor their lifestyle. The research and development of systems that provide a particular assistance to people is increasing in many areas of application. In particular, in the future, the recognition of ADL will be an important element for the development of a personal digital life coach, providing assistance to different types of users. To support the recognition of ADL, the surrounding environments should be also recognized to increase the reliability of these systems. The main focus of this Thesis is the research on methods for the fusion and classification of the data acquired by the sensors available in off-the-shelf mobile devices in order to recognize ADL in almost real-time, taking into account the large diversity of the capabilities and characteristics of the mobile devices available in the market. In order to achieve this objective, this Thesis started with the review of the existing methods and technologies to define the architecture and modules of the method for the identification of ADL. With this review and based on the knowledge acquired about the sensors available in off-the-shelf mobile devices, a set of tasks that may be reliably identified was defined as a basis for the remaining research and development to be carried out in this Thesis. This review also identified the main stages for the development of a new method for the identification of the ADL using the sensors available in off-the-shelf mobile devices; these stages are data acquisition, data processing, data cleaning, data imputation, feature extraction, data fusion and artificial intelligence. One of the challenges is related to the different types of data acquired from the different sensors, but other challenges were found, including the presence of environmental noise, the positioning of the mobile device during the daily activities, the limited capabilities of the mobile devices and others. Based on the acquired data, the processing was performed, implementing data cleaning and feature extraction methods, in order to define a new framework for the recognition of ADL. The data imputation methods were not applied, because at this stage of the research their implementation does not have influence in the results of the identification of the ADL and environments, as the features are extracted from a set of data acquired during a defined time interval and there are no missing values during this stage. The joint selection of the set of usable sensors and the identifiable set of tasks will then allow the development of a framework that, considering multi-sensor data fusion technologies and context awareness, in coordination with other information available from the user context, such as his/her agenda and the time of the day, will allow to establish a profile of the tasks that the user performs in a regular activity day. The classification method and the algorithm for the fusion of the features for the recognition of ADL and its environments needs to be deployed in a machine with some computational power, while the mobile device that will use the created framework, can perform the identification of the ADL using a much less computational power. Based on the results reported in the literature, the method chosen for the recognition of the ADL is composed by three variants of Artificial Neural Networks (ANN), including simple Multilayer Perceptron (MLP) networks, Feedforward Neural Networks (FNN) with Backpropagation, and Deep Neural Networks (DNN). Data acquisition can be performed with standard methods. After the acquisition, the data must be processed at the data processing stage, which includes data cleaning and feature extraction methods. The data cleaning method used for motion and magnetic sensors is the low pass filter, in order to reduce the noise acquired; but for the acoustic data, the Fast Fourier Transform (FFT) was applied to extract the different frequencies. When the data is clean, several features are then extracted based on the types of sensors used, including the mean, standard deviation, variance, maximum value, minimum value and median of raw data acquired from the motion and magnetic sensors; the mean, standard deviation, variance and median of the maximum peaks calculated with the raw data acquired from the motion and magnetic sensors; the five greatest distances between the maximum peaks calculated with the raw data acquired from the motion and magnetic sensors; the mean, standard deviation, variance, median and 26 Mel- Frequency Cepstral Coefficients (MFCC) of the frequencies obtained with FFT based on the raw data acquired from the microphone data; and the distance travelled calculated with the data acquired from the GPS receiver. After the extraction of the features, these will be grouped in different datasets for the application of the ANN methods and to discover the method and dataset that reports better results. The classification stage was incrementally developed, starting with the identification of the most common ADL (i.e., walking, running, going upstairs, going downstairs and standing activities) with motion and magnetic sensors. Next, the environments were identified with acoustic data, i.e., bedroom, bar, classroom, gym, kitchen, living room, hall, street and library. After the environments are recognized, and based on the different sets of sensors commonly available in the mobile devices, the data acquired from the motion and magnetic sensors were combined with the recognized environment in order to differentiate some activities without motion, i.e., sleeping and watching TV. The number of recognized activities in this stage was increased with the use of the distance travelled, extracted from the GPS receiver data, allowing also to recognize the driving activity. After the implementation of the three classification methods with different numbers of iterations, datasets and remaining configurations in a machine with high processing capabilities, the reported results proved that the best method for the recognition of the most common ADL and activities without motion is the DNN method, but the best method for the recognition of environments is the FNN method with Backpropagation. Depending on the number of sensors used, this implementation reports a mean accuracy between 85.89% and 89.51% for the recognition of the most common ADL, equals to 86.50% for the recognition of environments, and equals to 100% for the recognition of activities without motion, reporting an overall accuracy between 85.89% and 92.00%. The last stage of this research work was the implementation of the structured framework for the mobile devices, verifying that the FNN method requires a high processing power for the recognition of environments and the results reported with the mobile application are lower than the results reported with the machine with high processing capabilities used. Thus, the DNN method was also implemented for the recognition of the environments with the mobile devices. Finally, the results reported with the mobile devices show an accuracy between 86.39% and 89.15% for the recognition of the most common ADL, equal to 45.68% for the recognition of environments, and equal to 100% for the recognition of activities without motion, reporting an overall accuracy between 58.02% and 89.15%. Compared with the literature, the results returned by the implemented framework show only a residual improvement. However, the results reported in this research work comprehend the identification of more ADL than the ones described in other studies. The improvement in the recognition of ADL based on the mean of the accuracies is equal to 2.93%, but the maximum number of ADL and environments previously recognized was 13, while the number of ADL and environments recognized with the framework resulting from this research is 16. In conclusion, the framework developed has a mean improvement of 2.93% in the accuracy of the recognition for a larger number of ADL and environments than previously reported. In the future, the achievements reported by this PhD research may be considered as a start point of the development of a personal digital life coach, but the number of ADL and environments recognized by the framework should be increased and the experiments should be performed with different types of devices (i.e., smartphones and smartwatches), and the data imputation and other machine learning methods should be explored in order to attempt to increase the reliability of the framework for the recognition of ADL and its environments.
|
|---|