An Approach On Action Recognition-based Emotion Extraction

Action recognition based on motion data or video is the frontier of computervision and pattern recognition. However, given the result of action recognition, how toextract emotion information becomes a new cross subject of kinesiology and patternrecognition.Firstly, this research starts at the exploration of human emotion dataset.Combining kinesiology and computer science, a standard human emotion datasetcould be build. There are totally six emotions in the dataset, four basic emotions andtwo derived emotions. The basic emotions contain happiness, anger, sadness, fear andthe derived emotions are surprise and disgust. We think these six emotions are relativeto human activation. After we analyze human emotions and human action, we mapemotions to action, called action-emotion model.Secondly, considering our analysis of emotion model and emotion parameters,we put forward a new approach for emotion extraction. In this approach, the emotionsare divided by period which is the criterion of action division. The parameters areassigned weight respectively and summarized. The result is the emotion informationwe extracted.Finally, the approach has been extended to the platform based on video input.Because it is impossible for us to get the coordinates of each human bone directly, theextraction of bone parameters is replaced by feature extraction. Then we modify theapproach for adapting the input data and summarize them up. At last, we get the finalresult of emotion extraction.In experiment, it is proved that our approach performs well on the platform basedon human action data. The reason is everything on it is all coordinates and we couldget them or use them simply. On the other platform, our experiment shows that ourapproach has much room for improvement and a lot work to do in the future.