چكيده به لاتين
Nowadays, with the presence of CCTV cameras, people's actions can be monitored in public places. In addition, with the advent of smart homes, people's actions in their homes can be analyzed to achieve various goals such as recognizing emotions and also prevent emergency calls if necessary. From endangering people's health. Knowing people's action information is all the information we need to recognize people's feelings. In psychology, there are six basic emotions: happiness, anger, hatred, surprise, sadness and fear. The focus of this dissertation is on the three negative feelings of anger, fear and sadness that pose the greatest threat to the individual and society. To analyze emotion automatically, different types of input (text, audio, image and sensor data) are collected and analyzed from individuals. Manual methods were highly accurate but cost researchers a lot of time and money and included little research coverage. While automated methods automatically covered a large number of people in the research in question, they had lower financial and time costs compared to manual methods but were less accurate. In this dissertation, we present an automatic emotion detection method that accurately predicts (compared to previous methods) the three emotions of anger, fear and sadness. Our most important innovation in this research is the selection of the most important and distinctive frames in the construction of the proposed model, which leads to better learning and also prevents model saturation due to the large amount of data to teach the model. The proposed method includes the use of Mask R-CNN algorithm for human detection as well as convolution, two-dimensional deep learning for classification, and two algorithms, Detectron2 and VideoPose3D, which have been effective in selecting the most important and distinguishing features. For this purpose, we first use the Mask R-CNN to identify humans and calculate the ratio of the human body, and VideoPose3D and Detectron2 to create a three-dimensional graph (motion model) of the human body. Finally, we use a two-dimensional convolution classifier to predict three feelings of anger, fear and sadness. We use. Our results show that this method, on average, according to the four criteria of accuracy, F-measure, recall and precision 68.2, 75.1, 26.0 and 91.2 in terms of percentage compared to competing methods, in predicting the three feelings of anger, fear and sadness in humans It performs better.
