Engagement Estimation In Human-robot Interaction

With the influx of nuclear families, and the falling birthrate, the growingenvironment of children is beginning to change. Children in nuclear families areexposed to fewer companions as they would if in extended families. The robotcompanion is developed to satisfy this need.A series of problems about the playmate robot with regards to the children isdiscussed in this thesis. To act as a real friend of a child, the playmate robot shoulddetermine whether the child is engaged in the current game, and by which adjust itsbehavior automatically. This thesis focuses on how to estimate the child’s engagementwhen he or she is interacting with the playmate robot. The content of this thesis is asfollows:1. By introducing the design philosophy of playmate robot system and “child’smental state transition model”, we draw this thesis’s goal, which is to estimate user’sengagement state.2.3D video is acquired by using Kinect sensors. After the feature extraction,normalization and labeling, the samples are acquired.3. Gaussian mixture engagement model based3D data is proposed, and comparedto the Gaussian model based on2D data for engagement estimation. By experimentalevaluation, we prove that3D data’s performance is better than2D data. Additionally,the SVM for multi-class data set is implemented for this engagement estimationproblem.4. By analyzing the difference of evaluation methods accuracy and AUC, we areable to deduce the most appropriate method AUC to evaluate our experiments.5. Optimization for the preprocessing algorithm is proposed by analyzingassessment of engagement. After this improvement, AUC increases to0.7380from0.5935, which proves the efficiency of our algorithm optimization.