A Study Of Human Motion Recognition Algorithm Based On Wearable Sensor

Human motion contains abundant body language information and emotional information.User could get more natural and more efficient interaction experience with human motion.However,human motion is highly complex and diverse.With more precise need of motion capture,it is now the trend to use wearable sensors with multiple nodes.The increase of sensor nodes means that the dimension of motion capture data is rising.How to extract low dimensional and effective feature from motion capture data has become a challenge.Thus,the thesis focused on the motion feature.Aiming at the diversity of human motion,the thesis presented a quantitative human motion system,which represent human motion by few motion words.In order to eliminate redundant information in motion capture data,the thesis extract 6 meaningful angle features to describe each frame in motion sequence.After that,key pose extraction and hierarchical clustering are applied to angle features to generate motion words.Then each frame is replaced by motion word to represent the most similar pose,which reduces the diversity among human motions.The establishment of quantitative human motion system lays a key foundation for feature reduction and topic modeling.As traditional dimensionality reduction methods are inefficient for motion feature,the thesis applied generalized discriminant analysis to motion feature.The method maps motion feature vectors into high-dimensional features space through kernel function.In the transformed space,it's easy to solve the problem by linear discriminant analysis.In order to improve the performance of motion recognition,the thesis presented a novel motion descriptor: histogram of motion topic.The descriptor is based on semantic information discovered by topic model.As topic model ignores the grammatical information of motion,the thesis improved the motion descriptor by combining topic model with hidden Markov model,which is called histogram of topic-state.The thesis also designed a hierarchical support vector machine based on motion characteristic and application scenarios.Experiments show that generalized discriminant analysis has better ability to extract important information from motion sequence than other linear dimensionality reduction methods.The motion recognition accuracy for histogram of topic-state descriptor is 4% higher than sequence of the most informative joints descriptor,and has lower computational complexity than that.Compared to histogram of motion topic descriptor,the motion recognition accuracy for histogram of topic-state descriptor has increased 1% with little computational complexity sacrificed.