Fear classification using affective computing with physiological information and smart-wearables

Resumen

Among the 17 Sustainable Development Goals proposed within the 2030 Agenda and adopted by all United Nations member states, the fifth SDG is a call for action to turn gender equality into a fundamental human right effectively and an essential foundation for a better world. It includes the eradication of all types of violence against women. Focusing on the technological perspective, the range of available solutions intended to prevent this social problem is very limited. Moreover, most of the solutions are based on a panic button approach, leaving aside the usage and integration of current state-of-the-art technologies, such as the Internet of Things (IoT), affective computing, cyber-physical systems, and smart sensors. Thus, the main purpose of this research is to provide new insight into the design and development of tools to prevent and combat Gender-based Violence risky situations and, even, aggressions, from a technological perspective, but without leaving aside the different sociological considerations directly related to the problem. To achieve such an objective, we rely on the application of affective computing from a realist point of view, i.e. targeting the generation of systems and tools capable of being implemented and used nowadays or within an achievable time frame. This pragmatic vision is channelled through 1) an exhaustive study of the existing technological tools and mechanisms oriented to the fight Gender-based Violence, 2) the proposal of a new smart-wearable system intended to deal with some of the current technological encountered limitations, 3) a novel fear-related emotion classification approach to disentangle the relation between emotions and physiology, and 4) the definition and release of a new multi-modal dataset for emotion recognition in women. Firstly, different fear classification systems using a reduced set of physiological signals are explored and designed. This is done by employing open datasets together with the combination of time, frequency and non-linear domain techniques. This design process is encompassed by trade-offs between both physiological considerations and embedded capabilities. The latter is of paramount importance due to the edge-computing focus of this research. Two results are highlighted in this first task, the designed fear classification system that employed the DEAP dataset data and achieved an AUC of 81.60% and a Gmean of 81.55% on average for a subject-independent approach, and only two physiological signals; and the designed fear classification system that employed the MAHNOB dataset data achieving an AUC of 86.00% and a Gmean of 73.78% on average for a subject-independent approach, only three physiological signals, and a Leave-One-Subject-Out configuration. A detailed comparison with other emotion recognition systems proposed in the literature is presented, which proves that the obtained metrics are in line with the state-of-the-art. Secondly, Bindi is presented. This is an end-to-end autonomous multimodal system leveraging affective IoT throughout auditory and physiological commercial off-the-shelf smart-sensors, hierarchical multisensorial fusion, and secured server architecture to combat Gender-based Violence by automatically detecting risky situations based on a multimodal intelligence engine and then triggering a protection protocol. Specifically, this research is focused on the hardware and software design of one of the two edge-computing devices within Bindi. This is a bracelet integrating three physiological sensors, actuators, power monitoring integrated chips, and a System-On-Chip with wireless capabilities. Within this context, different embedded design space explorations are presented: embedded filtering evaluation, online physiological signal quality assessment, feature extraction, and power consumption analysis. The reported results in all these processes are successfully validated and, for some of them, even compared against physiological standard measurement equipment. Amongst the different obtained results regarding the embedded design and implementation within the bracelet of Bindi, it should be highlighted that its low power consumption provides a battery life to be approximately 40 hours when using a 500 mAh battery. Finally, the particularities of our use case and the scarcity of open multimodal datasets dealing with emotional immersive technology, labelling methodology considering the gender perspective, balanced stimuli distribution regarding the target emotions, and recovery processes based on the physiological signals of the volunteers to quantify and isolate the emotional activation between stimuli, led us to the definition and elaboration of Women and Emotion Multi-modal Affective Computing (WEMAC) dataset. This is a multimodal dataset in which 104 women who never experienced Gender-based Violence performed different emotion-related stimuli visualisations in a laboratory environment. The previous fear binary classification systems were improved and applied to this novel multimodal dataset. For instance, the proposed multimodal fear recognition system using this dataset reports up to 60.20% and 67.59% for ACC and F1-score, respectively. These values represent a competitive result in comparison with the state-of-the-art that deals with similar multi-modal use cases. In general, this PhD thesis has opened a new research line within the research group under which it has been developed. Moreover, this work has established a solid base from which to expand knowledge and continue research targeting the generation of both mechanisms to help vulnerable groups and socially oriented technology.