Research On Improved Algorithm For Sparse Representation Of Motion Recognition

Sparse representation(SR)is an efficient technology of processing signal.It is consistent with the principle of human perceiving the visual information,which has a good performance in many areas of computer vision.Among them,human action recognition(AR)in videos has been widely applied in intelligent monitoring,human-computer interaction and multimedia analysis.Therefore,the study of AR in videos based on SR attracts lots of attention.There are many feature extraction methods for videos.In theory,combining multiple features should improve the performance of AR.However,the simple fusion methods can’t achieve the satisfying result because the distribution in feature space of different channels is quite different.In addition,the contribution of each feature and dictionary atom is uneven in traditional SR for classification.We can’t make full use of their information if they are equally treated.To solve these problems in AR,we proposed the improved SR methods by combining RGB-D features and mining the discriminativeness of elements.1)To improve the effectiveness of combining multiple features,we proposed the improved minimum reconstruction error and maximum number of non-zero items for fusing RGB-D features.For a kind of feature,we firstly learned a class-specific dictionary for each class.Given a test sample,we obtained its sparse coefficients corresponding to the sub-dictionary.Finally,we recognized an action by calculating the fusion reconstruction error and the fusion number of non-zero items.These methods can combine RGB-D features effectively in AR and analyze the contribution of each feature.2)We proposed the discriminative-element-aware SR by mining the discriminativeness of features and dictionary atoms.Firstly,we preprocessed the training data by utilizing the discriminativeness of features,then we computed the discriminativeness of dictionary atoms which has been learned by the preprocessed training data.Given a test video,we computed the discriminativeness of its features,and finally classified this video by the discriminativeness of its features and dictionary atoms.This method is less sensitive to noise.The experiments have proved that the proposed methods can improve the performance of AR.