Taijiquan Trajectory Segmentation And Recognition Based On Multi-sensor Data Fusion

As modern science and technology are making progress rapidly,the technology about action recognition has been widely used in many fields,such as interaction between human and computer,medical treatment,sports,security monitoring and so on.Taijiquan is not only a way of health care,but also a national certified sport competition.Considering that the use of action recognition technology can effectively improve the level of athletes,so this paper designs and implements the acquisition and recognition system of Taijiquan action.This paper will use inertial sensors and visual sensors to collect the movement data of Taijiquan,and through the analysis and processing of the data,finally complete the automatic segmentation and recognition classification of Taijiquan action.First of all,this paper separately designs the Taijiquan motion capture system based on inertial sensor system and vision sensor system.Each system can independently collect the Taijiquan motion data.We use the human joint motion trajectory in three-dimensional space to represent the Taijiquan action.Because of the noise interference,the trajectory which is directly obtained by the sensor needs to be smoothed.The data obtained by the vision sensor can get better results after guided filtering and Kalman smoothing.The measurement result of inertial sensor is smoothed by moving average in the mode with five points.Finally,according to the instrument characteristics of the two sensors,an inertial-visual joint sampling system based on multi-sensor data fusion is designed.Secondly,according to the principle of minimum description length,the appropriate feature points are selected as the segmentation points to segment the action trajectory.We transform the sub trajectory into vectors,and the starting points of all vectors are changed into the origin to complete the vectorization of the trajectory.The time length corresponding to the sub trajectories and the vectorized sub trajectories constitute a four-dimensional feature vector.DBSCAN is used to cluster the feature vectors,and the optimal clustering results are selected according to a variety of clustering evaluation indexes.Finally,according to the clustering results,an automatic segmentation method of Taijiquan trajectory based on dynamic time warping is designed.This method can judge the ending position of the move according to the template provided by multiple joint points,and the better automatic segmentation effect can be obtained by segmenting the ending position.For the segmented trajectory,this paper designs two kinds of classification and recognition methods: the first is the recognition algorithm for single pass nodes.Firstly,the clustering results obtained in the previous part are transformed into the form of Gaussian mixture model,and then into the standard Fisher vector.Fisher vector is put into support vector machine to train classifier.The second is the recognition algorithm for multi joint points.This algorithm still uses DWT to calculate the similarity between trajectories,and uses template matching method to complete the action recognition.In the experiment of Chapter 5,the accuracy of the classifier under different conditions is given.