Brain computer interface for analyzing emotions

Emotion is the subjective experience that reflects our mental states and can significantly affect our cognitive function and action tendencies. With the advances in artificial intelligence (AI) and brain-computer interface (BCI) technologies, the ability for computer applications to recognize human emotions can provide us more intelligent services, such as style-adjusting e-learning system , driver's fatigue detection, e-healthcare assistance, etc. To recognize human emotions using machines, scientists studied many different sources of signal. Literature shows that electroencephalograph (EEG) demonstrates promising characteristics in revealing high level brain activities, and its correlation with emotions is also profound. This made EEG-based emotion recognition a hot research area under BCI. However, the majority approaches have problems in achieving reliable recognition accuracies. Specifically, we face the following challenges: (i) the features extracted from EEG may not be good enough for recognizing emotional activities; (ii) EEG signal is extremely susceptible to noise and artifact interference, which may affect the recognition accuracy; and (iii) increasing feature dimension may enhance the recognition accuracy, but as a cost, the whole process will incur much more computation time. In this project, I investigated different features extraction methods from the literature and explored different techniques to process those features. I also compared the classification performance on different classifiers. To address the three challenges aforementioned, I propose a fast and robust EEG-based emotion recognition model that applies feature smoothing on a set of statistical features extracted from the raw EEG signals. I performed experiments on benchmarks affective EEG datasets and obtained convincing experimental results. The main contributions of this project are listed as follows. Firstly, I had a comprehensive study on related work, and identified the features that are suitable for emotion recognition and computationally simple to extract. Secondly, I propose a novel feature smoothing method that can enhance the quality of features and provided the theoretical proof on the effectiveness of this method. Lastly, the EEG-based emotion recognition model I proposed outperforms prior studies on benchmark affective EEG datasets. Compared to other studies on the DEAP and SEED datasets, my method kept the shortest processing time for features and achieved the highest classification accuracy.Bachelor of Engineering (Computer Science