Research On Human Behavior Detection And Recognition Based On Feature Attributes

With the rapid development of video surveillance,smart home and smart driving,human behavior detection and identification has shown great market demand and great scientific research value.In practical applications,the shooting angle of the video,the background light,the variability of the human body characteristics and the relationship between the human body behaviors,etc.,make the human behavior detection and recognition based on the feature attributes become the focus of researchers.This thesis proposes an accidental action pedestrian detection method based on feature extraction.It solves the problem of the prior art due to the problem that the pedestrian's limb movement is too large,such as bending and squatting,which makes up for the deficiency of the prior art.Test and analysis were carried out,and other methods were used to compare the experiments again,which made the effectiveness of the method fully proved.A tensor-based graph matching behavior recognition method is proposed.Considering the problem of space and time complexity,the features are extracted from the original video and processed to form higher-order feature descriptors.The extracted high-order features are used.Learning training,classification by learning algorithm and testing on public behavior data,the experimental results verify the effectiveness of the method.A behavior recognition method based on feature learning graph matching is proposed.The key of this method is to learn the class model map to complete the matching recognition.Combining with structured support vector machine and middle-level SIFT features,the experiments were carried out on KTH and UCF50 behavior datasets,and different algorithms were selected for experimental analysis and comparison.The statistical results show that the recognition performance of this method is outstanding,and the effectiveness of the behavior recognition method based on feature learning graph matching is verified.