الفهرس | Only 14 pages are availabe for public view |
Abstract ObjectivesofStudy The detection of human activities with high accuracy is our mission. the main objective is to design and develop a HAR-IoT system of appropriate quality to detect online human activities via internet of things. We need to achieve the following sub-goals in order to reach the main goal: 1. We needed to obtain an overview of previous work carried out in the field of human activity recognition and to get an understanding of the techniques underlying such systems in order to construct a HAR-IoT system of appropriate quality. Both chapters, background and literature review, explain this point. 2. Build classification model, decided to be DL. 3. Enhance the architecture of deep learning techniques. This point has been clarified in methodology chapter. 4. For measuring the efficiency of third goal, proposed method will be compared with RNN and CNN, with two different architectures, as well as three different dataset were used based on dependent and independent approaches. This point has been clarified in experiments chapter. 5. Collected our dataset to test on real data. Recently,InternetofThings(IoT)ismoresignificantasitisbasedoninternetwhichisusedin our daily life. It is beneficial to exploit the potential of IoT for developing Human Activities Recognition (HAR) systems. HAR is a wide field of study concerned with distinguishing particularactivitiesofahuman. Themostofthisactivitiesareindoorssuchasdown/upstairs, walking, standing, lying, sitting. Indeed, HAR is a vital research problem as it is utilized in different areas. Therefore this work focuses on gathering a real-time dataset from the proposed IoT system to track and record the data of human activities. For improving the classification technique, we suggested a deep learning method to enhance the architecture of ConvolutionNeuralNetwork(CNN)byusinggeneticalgorithm. Thereforegeneticalgorithm (GA) is utilized as an enhancing method to get the optimal value parameters of the deep learning. In the flatten layer, we combined the features that are extracted automatically from CNN with handcrafted features, which include the global context of data. Two public datasets (UCI and WISDM) were used to train the proposed model. In order to assess the effectiveness of the proposed method, the live dataset that has been collected by our IoT system is used. Our model improved accuracy up to 93.8% and 86.1% for user-dependent and independent approaches. |