Research On Action Recognition Based On Spatial-temporal Features Of Trajectories

Human action recognition is one of the most important research directions in machine learning,it has achieved rapid development both in academic research and practical application.At present,human action recognition and video classification technologies have been widely used in many intelligent industries,and they have a broad market space and future.However,advanced technologies are often come together with a lot of data.Efficient analysis and processing of a large number of videos from real life is the goal of researching video analytics technologies.The research on advanced methods plays a crucial role in the promotion of action recognition.Although we have made some progress under the tireless efforts of our predecessors,we still face many challenges.The feature representation of action is the basis for the study of action recognition methods,therefore,this thesis takes the exploration of more effective features as a starting point and conducts the following research.First of all,this thesis summarizes the latest popular feature construction methods in the study of human action recognition at home and abroad,and analyzes the defects that still exist in existing features.In view of the large number of traditional dense trajectories and there is redundancy,in this thesis,only the more important trajectories are selected for feature construction by calculating the discrete coefficients of the trajectories.In addition,inspired by the importance of the relationship between speed and acceleration for definition of motion in physics,we have co-occurrence statistics of speed and acceleration to construct the velocity and acceleration co-occurrence descriptors,and the movement trend information was mined to describe the movement of the trajectory more accurately.Secondly,this thesis considers that the traditional trajectory-based features take the trajectory as an independent individual and ignores the neighbor information of trajectory.Therefore,we derive the neighbors of trajectory by KNN algorithm.In order to calculate rich neighborhood information around the trajectory,this thesis calculates the complex relationship between the center trajectory and the neighboring trajectory from four different angles including absolute motion,relative motion,distance relation,and directional relation.Then,we use the nine typical measures that describe the distribution characteristics of data in statistics including mean,median,maximum,minimum,range,variance,dispersion coefficient,skewness,and kurtosis to describe the distribution information of the neighbors between trajectories,and the spatial-temporal neighbor distribution descriptors between trajectories are constructed.This thesis starts from four different perspectives and can complement each other,improving the ability of the feature to describe complex and varied behaviors.Thirdly,for the existing coding methods in the coding process,the problem of fuzzy boundary between different clusters in the same category and similar clusters in different categories is ignored,that make coding process introduce confusion.To solve this problem,the large margin nearest neighbor coding method is proposed in this thesis.By learning the new distance measurement,the samples that are easy to be confused in the transformation space are separated as much as possible,and the accuracy of projection is improved by taking advantage of the algorithm itself.Further,the confusion introduced in the coding process is reduced and the coding results are optimized.In addition,the information from the four angles of the spatial and temporal nearest neighbor distribution descriptor is further encoded by learning different transformation matrices through the encoding method proposed in this thesis and the coding result is used for action recognition after combination,and the recognition result is further optimized.Finally,we summarize the main research content of this thesis,and give future research directions.