|dc.description.abstract||Recognizing human activities in domestic environments from audio and active power consumption sensors is a challenging task since on the one hand, environmental sound signals are multi-source, heterogeneous, and varying in time and on the other hand, the active power consumption varies significantly for similar type electrical appliances.
Many systems have been proposed to process environmental sound signals for event detection in ambient assisted living applications. Typically, these systems use feature extraction, selection, and classification. However, despite major advances, several important questions remain unanswered, especially in real-world settings. A part of this thesis contributes to the body of knowledge in the field by addressing the following problems for ambient sounds recorded in various real-world kitchen environments: 1) which features, and which classifiers are most suitable in the presence of background noise? 2) what is the effect of signal duration on recognition accuracy? 3) how do the SNR and the distance between the microphone and the audio source affect the recognition accuracy in an environment in which the system was not trained? We show that for systems that use traditional classifiers, it is beneficial to combine gammatone frequency cepstral coefficients and discrete wavelet transform coefficients and to use a gradient boosting classifier. For systems based on deep learning, we consider 1D and 2D CNN using mel-spectrogram energies and mel-spectrograms images, as inputs, respectively and show that the 2D CNN outperforms the 1D CNN. We obtained competitive classification results for two such systems and validated the performance of our algorithms on public datasets (Google Brain/TensorFlow Speech Recognition Challenge and the 2017 Detection and Classification of Acoustic Scenes and Events Challenge).
Regarding the problem of the energy-based human activity recognition in a household environment, machine learning techniques to infer the state of household appliances from their energy consumption data are applied and rule-based scenarios that exploit these states to detect human activity are used. Since most activities within a house are related with the operation of an electrical appliance, this unimodal approach has a significant advantage using inexpensive smart plugs and smart meters for each appliance. This part of the thesis proposes the use of unobtrusive and easy-install tools (smart plugs) for data collection and a decision engine that combines energy signal classification using dominant classifiers (compared in advanced with grid search) and a probabilistic measure for appliance usage. It helps preserving the privacy of the resident, since all the activities are stored in a local database.
DNNs received great research interest in the field of computer vision. In this thesis we adapted different architectures for the problem of human activity recognition. We analyze the quality of the extracted features, and more specifically how model architectures and parameters affect the ability of the automatically extracted features from DNNs to separate activity classes in the final feature space. Additionally, the architectures that we applied for our main problem were also applied to text classification in which we consider the input text as an image and apply 2D CNNs to learn the local and global semantics of the sentences from the variations of the visual patterns of words. This work helps as a first step of creating a dialogue agent that would not require any natural language preprocessing.
Finally, since in many domestic environments human speech is present with other environmental sounds, we developed a Convolutional Recurrent Neural Network, to separate the sound sources and applied novel post-processing filters, in order to have an end-to-end noise robust system. Our algorithm ranked first in the Apollo-11 Fearless Steps Challenge.||en