Research On Behavior Recognition Method Based On Human And Object Interaction

The rapid development of human behavior recognition technology has gradually penetrated into many industries.It is an important step for computers to automatically understand real scenes.It has broad application prospects,and more and more researchers are investing in it.By analyzing the complexity of human behavior,it is found that compared to single-person behavior,the recognition of interactive behaviors has high complexity,a large amount of redundant information,high feature dimensions and difficult to distinguish.Therefore,the characteristics of interactive behaviors are selected and expressed It plays a vital role in the recognition algorithm.Most of the researches on interactive actions are based on RGB images.Although good research results have been obtained,there are some limitations.For RGB images,complex background,light intensity,angle and other factors will affect the interaction behavior recognition and reduce the recognition rate.Therefore,in this paper,we mainly focus on how to select feature extraction of interactive actions in combination with depth images and optimize algorithm efficiency.The research work of this paper is as follows:(1)Using the complementarity of multi-source information,the respective features are extracted accordingly.This paper first uses Canny operator to extract edge features for depth images,and uses Local Binary Pattern operator rotation invariant mode to extract texture features for RGB images,and uses Histograms of Oriented Optical Flow to describe dynamic features;then perform weighted fusion on the extracted edge features and texture features;then use sparse coding space pyramid matching model to encode the static fusion features and optical flow motion trajectory features into a pool,and the resulting pool Then the features are fused again.Finally,Support Vector Machine(SVM)is used for classification to realize the interaction behavior recognition.For the contribution of features,a weight fusion mechanism is used to select the best feature fusion.(2)The complexity of the interaction is high,and the amount of information redundancy is large.More powerful features need to be extracted.To solve this problem,this paper uses a low-rank sparse optimization algorithm to remove the interference of redundancy and noise to the maximum extent,and extract a clean Low rank matrix.In this paper,Robust Principal Component Analysis is used to extract the low-rank matrices fused with the coding features.Finally,the low-rank matrices of the features are classified and identified using the SVM algorithm.For complex problems,a sparse super-complete dictionary method is used to compress the space,and a low-rank sparse optimization algorithm is used to improve the efficiency of the algorithm and reduce the complexity of the algorithm.In order to prove the effectiveness and superiority of the proposed algorithm,experiments were performed on several common behavioral datasets,including CAD-60 dataset,MSR Action Pairs dataset and SBU dataset.The first two data sets are used to identify the person's interaction,and the third data set is used to identify the two-person interaction.This paper compares each data set with other excellent algorithms in the literature.The experimental results show that the recognition results of the proposed algorithm have improved.