Sensor Applications for Human Activity Recognition in Smart Environments

AutorFu, Biying
BetreuerKuijper, Arjan
AbstraktHuman activity recognition (HAR) is the automated recognition of individual or group activities from sensor inputs. It deals with a wide range of application areas, such as for health care, assisting technologies, quantified-self and safety applications. HAR is the key to build human-centred applications and enables users to seamlessly and naturally interact with each other or with a smart environment. A smart environment is an instrumented room or space equipped with sensors and actuators to perceive the physical state or human activities within this space. The diversity of sensors makes it difficult to use the appropriate sensor to build specific applications. This work aims at presenting sensor-driven applications for human activity recognition in smart environments by using novel sensing categories beyond the existing sensor technologies commonly applied to these tasks. The intention is to improve the interaction for various sub-fields of human activities. Each application addresses the difficulties following the typical process pipeline for designing a smart environment application. At first, I survey most prominent research works with focus on sensor-driven categorization in the research domain of HAR to identify possible research gaps to position my work. I identify two use-cases: quantified-self and smart home applications. Quantified-self aims at self-tracking and self-knowledge through numbers. Common sensor technology for daily tracking of various aerobic endurance training activities, such as walking, running, or cycling are based on acceleration data with wearable. However, more stationary exercises, such as strength-based training or stretching are also important for a healthy life-style, as they improve body coordination and balance. These exercises are not well tracked by wearing only a single wearable sensor, as these activities rely on coordinated movement of the entire body. I leverage two sensing categories to design two portable mobile applications for remote sensing of these more stationary exercises of physical workout. Sensor-driven applications for smart home domain aim at building systems to make the life of the occupants safer and more convenient. In this thesis, I target at stationary applications to be integrated into the environment to allow a more natural interaction between the occupant and the smart environment. I propose two possible solutions to achieve this task. The first system is a surface acoustic based system which provides a sparse sensor setup to detect a basic set of activities of daily living including the investigation of minimalist sensor arrangement. The second application is a tag-free indoor positioning system. Indoor localization aims at providing location information to build intelligent services for smart homes. Accurate indoor position offers the basic context for high-level reasoning system to achieve more complex contexts. The floor-based localization system using electrostatic sensors is scalable to different room geometries due to its layout and modular composition. Finally, privacy with non-visual input is the main aspect for applications proposed in this thesis. In addition, this thesis addresses the issue of adaptivity from prototypes towards real-world applications. I identify the issues of data sparsity in the training data and data diversity in the real-world data. In order to solve the issue of data sparsity, I demonstrate the data augmentation strategy to be applied on time series to increase the amount of training data by generating synthetic data. Towards mitigating the inherent difference of the development dataset and the real-world scenarios, I further investigate several approaches including metric-based learning and fine-tuning. I explore these methods to finetune the trained model on limited amount of individual data with and without retrain the pre-trained inference model. Finally some examples are stated as how to deploy the offline model to online processing device with limited hardware resources. Personalization is the task that aims at improving quality of products and services by adapting itself to the current user. In the context of automotive applications, personalization is not only about how drivers sets up the position of their seat or their favorite radio channels. Going beyond that, personalization is also about the preference of driving styles and the individual behaviors in every maneuver executions. One key challenge in personalization is to be able to capture and understand the users from the historical data produced by the users. The data are usually presented in form of time series and in some cases, those time series can be remarkably long. Capturing and learning from such data poses a challenge for machine learning models. To deal with this problem, this thesis presents an approach that makes uses of recurrent neural networks to capture the time series of behavioral data of drivers and predict theirs lane change intentions. In comparison to previous works, our approach is capable of predicting not only driver's intention as predefined discrete classes (i. e. left, right and lane keeping) but also as continuous values of the time left until the drivers cross the lane markings. This provides additional information for advanced driver-assistance systems to decide when to warn drivers and when to intervene. There are two further aspects that need to be considered when developing a personalized assistance system: inter- and intra-personalization. The former refers to the differences between different users whereas the later indicates the changes in preferences in one user over time (i. e. the differences in driving styles when driving to work versus when being on a city sightseeing tour). In the scope of this thesis, both problems of inter- and intra-personalization are addressed and tackled. Our approach exploits the correlation in driving style between consecutively executed maneuvers to quickly derive the driver's current preferences. The introduced networks architecture outperforms non-personalized approaches in predicting the preference of driver when turning left. To tackle inter-personalization problems, this thesis makes use of the Siamese architecture with long short-term memory networks for identifying drivers based on vehicle dynamic information. The evaluation, which is carried out on real-world data set collected from 32 test drivers, shows that the network is able to identify unseen drivers. Further analysis on the trained network indicates that it identifies drivers by comparing their behaviors, especially the approaching and turning behaviors.