Research On Behavior Semantic Extraction Algorithms For Specific Scenarios

With the development of science and technology,the resolution of video images has gradually increased,and the hardware computing performance has also increased geometrically.This makes it possible for machine learning technology to break through the bottleneck of computing power and promote its wider application to all aspects of people's production and life.Behavior recognition and action classification currently have important research significance in security,medical care and other aspects.Effective use of machine learning technology can effectively promote the further development of behavior recognition and action classification.Human behavior recognition is currently mainly reflected in the classification of actions,and action classification can be seen as a combination of feature extraction and classifier design.However,the deformation of the video image,the change of viewing angle and illumination,and the relative movement of the camera position will increase the difficulty of feature extraction and classifier classification,and ultimately affect the recognition results.At the same time,the existing detection methods lack an overall analysis of the video sequence.In order to solve the above problems,this paper studies behavior recognition and action classification and introduces machine learning network models for experiments.The innovation of this study is mainly reflected in the following two aspects:First,a behavior recognition model that combines skeleton information and convolutional network is proposed,which can better adapt to the problems of uneven training samples and varying regional scales,reducing the impact of video image deformation and lighting.The rapid development of high-resolution video images has resulted in the increase of redundant information.The input of the skeleton network is the extracted three-dimensional skeleton joint information.The displacement of the joint points is used to characterize the movement of the human limbs,which effectively reduces the video.The impact of redundant information.This model also innovatively introduces the attention mechanism.A novel spatio-temporal convolution module is designed in the network to model the spatio-temporal correlation of human skeleton motion,which can better perform the overall video sequence.Analysis improves the effectiveness of training and the accuracy of classification.Secondly,this study sorts,analyzes and selects the existing data sets.At the same time,in order to avoid the change of the background in the image,a data collection scene was built in the laboratory to simulate a relatively fixed background environment.Based on this specific environment,a new data set is constructed,and skeleton features are extracted from the video clip samples in the data set.