The Study Of Behavior Recognition Based On RGB Image And Depth Information

Behavior recognition has a very broad application prospects in intelligent monitoring system,intelligent transportation systems, long-distance video conference,human-computer interaction, medical diagnosis,video retrieval, and so on,so human behavior recognition become one of the most important research topic in computer vision field in recent years.Behavior recognition using image processing,pattern recognition and other methods,makes the computers understand behaviors in a video sequence and then automatically identify a variety of common behaviors.In this paper, we propose a new human behavior method,which fuses two features in the same pattern together. The experimental evaluations show that the proposed approach has achieved high recognition accuracy.what's more,we design a traffic gestures recognition system based on skeleton joint after analyzing the computational complexity and time cost about features fusion methods.At present,most behavior recognition methods based on single feature.We propose a method based on RGB-D features fusion.Firstly, HOG3D descriptor describes space-time interest points and HOG3D feature is extracted on RGB images.Then,We construct 4D normal vector histogram to describe depth information based on point cloud surface normal vector from 3D surface point cloud information in depth images.Finally we use Canonical Correlation Analysis(CCA) to perform feature dimensional reduction and fusion. And Support Vector Machine(SVM) is explored for model testing and classifying for the new features.We design a traffic gestures recognition system based on skeleton joint after analyzing the computational complexity and time cost about features fusion methods. Distance weighting dynamic time warping (DW-DTW) and K-nearest neighbor algorithm are used to recognize the real time traffic gestures on pioneer and move in the right direction.