Activity Analysis Based On Splitting Structure With Temporal Correlation

Complex activity recognition is an important component of video analysis.The complex activity often contains several actionlets,which are interrelated with each other.actionlets are important parts of complex activity,and can interpret and analyze activity process.Thus it has widely application in the surveillance video analysis,human-computer interaction and video content retrieval task.Though existing models could achieve promising performance for simple activity analysis and recognition,it still exist many limitations in the complex activity analysis:(1)currently complex activity recognition is mainly dependent on low-level features,thus difficult to explain activity process;(2)simple actionlets are difficult to be defined clearly;(3)there are more redundant background information between simple actionlets in complex activity than simple activity.In this thesis,based on the analysis of the activity task,Combined with important characteristics of varieties of simple actionlets constitute in complex activity,we mainly focus on the following difficulties:firstly,it is difficult to define simple actionlets when it is lacking in groundtruth of actionlets;Secondly,how to learn reasonable temporal structure to explain the activity process is another problem;lastly,When there are more background and noise information in complex activity,it is hard to remove the noise and obtain more robust activity expression,used for the action localization task.Considering the above difficulties,the main work of this thesis is as follows:(1)As for how to learn reasonable temporal structure of activity to explain the process,we combine relative temporal structure of actionlets and Grenander pattern theory.By analyzing the independence of actionlets and related temporal structure of actionlets,we eliminate untrusted actionlets and combine closely linked actionlets based on pattern theory and reassemble temporal sequence of actionlets in the framework of probability;with maximizing probability of temporal structure,we can obtain the optimal temporal structure and realize activity recognition using dynamic time warping algorithm.(2)For the difficulty on studying actionlets expression of complex activity effectively,we learn actionlets expression by unsupervised way,mining the latent semantics of actionlets,thus achieving hierarchy analysis of activity.(3)To solve complete action localization in spite of noise and background information disturb analysis,the thesis learns the actionlets sequence and analyzes the temporal relationship between them modeling time-series action instance;the thesis achieves action instance localization completely by reassembling actionlets and analyzing their characteristics.