Research On Behavior Recognition Based On The Sparse Representation Model

Vision-based behavior recognition refers to the classification of a video or image sequence containing a certain behavior into the correct category label.Behavior recognition technology can be applied in many fields such as video surveillance,self-driving cars,human-computer interaction,sports video analysis and so on.It has a wide range of applications.Behavior recognition research covers many tasks such as feature extraction,behavior description and representation,classifier design and so on.It involves the intersection of computer vision,machine learning,pattern recognition and other disciplines.Behavior recognition is a hot topic in the field of artificial intelligence in recent years and has important research significance.Based on the comprehensive study of the relevant literature,the behavior recognition method based on sparse representation model is deeply researched in this dissertation.The representation of variations for the same class and differences between different classes is a key issue in behavior recognition.In recent years,the sparse representation based classifier(SRC)has been widely used in the field of pattern recognition.The SRC model can solve the problem above to a certain extent.In the sparse model,the dictionary is learned separately for each behavior category,so that the atoms in the dictionary contain different patterns as much as possible.The intra-class differences are included in the class specific dictionary.The dictionaries learned for different behaviors can reflect the differences between classes.This dissertation mainly researches the theory of sparse representation and builds a behavior recognition algorithm based on sparse representation.The main research contents and contributions of the dissertation are as follows:(1)A behavior recognition method based on the nearest neighbor atom representation of the local spatio-temporal features on the concatenated dictionary is proposed.The traditional SRC aims at minimizing the reconstruction error of local spatio-temporal features in the recognition stage.Excessive pursuit of the minimum of reconstruction error will lead to a wide choice of atomic distribution,which weakens the discriminative ability of the dictionary.To solve this problem,a recognition method based on nearest neighbor atom representation is proposed.The basic principle of the model is similar to that of KNN classifier.At the training phase,the proposed method uses the K-SVD algorithm to learn a separate dictionary for each category;in the recognition stage,the coefficients represented by the nearest atoms of the local features of test video in the concatenated dictionary are counted.The label is determined by the number of coefficients.Experiments on the Facial Expressions public dataset and the Weizmann human behavior dataset show that the proposed algorithm is better than the traditional SRC classifier.(2)A supervised dictionary learning method which combines the group sparse prior knowledge in the coding phase is proposed.The model uses the label information of each category in combination with the structure of the concatenated dictionary in the training phase to penalty different parts of the sparse coefficients with different values.In addition to learning the dictionary separately for each category,a common pattern dictionary is also learned to reduce the reconstruction error.This will promote the discriminability of each class specific dictionary to some extent.Algorithm for optimizing the proposed model is given and experiment was performed on Facial Expressions,Hand Gesture,and UCF Sports public datasets.The results show that the classification accuracy of the proposed method is improved when compared with other SRC models.(3)Learning a kernel dictionary linearly with group sparse priors is proposed.Learning a kernel space dictionary can further improve the classification accuracy of the SRC model.The dictionary is learned in a linearized manner by mapping the features to virtual features.A dictionary is learned by combing the kernel space learning methods and the group sparse prior for behavior recognition.