| Artificial intelligence and human–computer interactions are showing a trend of rapid growth,and emotion recognition plays an important role in improving the intelligence of human-computer interaction.Human motions play a key role in emphasizing and conveying emotions to meet the complexity of daily application scenarios.Exploring the hidden emotional states from human motion and make accurate identification,which helps to promote the development of non-verbal behavior in human-computer interaction and equip machines to have emotional communication capabilities.We consider that this is helpful to improve the intelligence and emotionalisation of human-computer interaction at this stage.Recognizing human emotional state from inertial motion data is an emerging research direction in recent years,it uses inertial sensors to obtain human motion data and uses machine learning algorithms to analyze and recognize emotional information.Therefore,this thesis studies some problems in human emotion recognition methods driven by inertial motion data.The main contents include:(1)A limb motion emotion recognition method based on emotion annotation is proposed.Firstly,this thesis manually annotates 8 kinds of limb movements and determines the corresponding relationship between limb movement and emotion through fuzzy comprehensive evaluation.Secondly,a lightweight convolutional neural network(LW-CNN)is designed to extract the critical and discriminative features.Thirdly,channel attention-graph convolutional network(CA-GCN)is used to aggregate and enhance the relevant features of multi-sensor signals.Finally,a weighted kernel support vector machine(WK-SVM)model is proposed to realize the flexible expression of heterogeneous data and emotion recognition.The experimental results show that the proposed limb motion emotion recognition method can effectively identify emotions from limb motion.(2)A walking emotion recognition method based on immersive virtual reality(VR)technology is proposed.The emotional simulation ability of the VR environment effectively stimulates real and profound emotions.Firstly,the walking motion emotion data is represented as an image by discrete wavelet transform(DWT)and used as the input of CNN to improve the system performance.Secondly,this thesis takes into account the feature level and decision level fusion of multi-inertial sensor data.The feature fusion module based on self-attention automatically learns and allocates the importance weights of each sensor,and the decision fusion module based on CRITIC method assigns weights to the prediction labels that may affect the final decision.The experimental results show that the proposed walking motion emotion recognition method can obtain emotional information from walking motion and accurately identify it.(3)An error correction algorithm for inertial motion data in human emotion recognition is proposed.Firstly,the inertial sensor is fixed in different directions and remains stationary to collect the reference data.Secondly,the rotation matrix is obtained by the principle of matrix rotation transformation and the error data is corrected.Finally,the time domain and frequency domain features of the inertial data are manually extracted and input into the SVM model to verify the effectiveness of the correction algorithm.The experimental results show that the proposed correction algorithm can effectively reduce the influence of inertial motion data error caused by inertial sensor position change on human emotion recognition. |
