Research On Key Techniques Of Human Action Recognition In Video Surveillance

As a key technology in intelligent monitoring,human action recognition is an active research topic in the field of computer vision.The research on single human action recognition has gained plentiful and substantial harvest.Due to the spatial complexity of human interaction,the differences in action characteristics at different time periods,and the complexity of interactive action features,the recognition of human interaction is more challenging.This paper studies the recognition of human interaction in video surveillance,and mainly makes the corresponding improvement in the establishment of human interaction recognition model,the difference in characteristics at different time periods and the complexity of human interaction features.The improvement measures are as follows:(1)In the aspect of feature extraction of human interaction,this paper proposes a multi-feature fusion network algorithm based on parallel Inception and ResNet.The Inception module uses different scales of feature receptive fields,reducing the amount of network parameters and improving network performance effectively.The ResNet module mitigates degradation problem caused by increased depth of neural networks and achieves higher classification accuracy.The migration learning method is used to extract feature by using Inception network and ResNet respectively,and then the extracted features are merged and training is continued to realize parallel connection of multi-feature neural networks.In this paper,experiments on UT datasets show that the multi-model fusion convolutional neural network has higher action recognition accuracy than the original single feature extraction network.(2)Considering the characteristics of human-human interaction,this paper proposes a method of human interaction recognition based on the whole-individual segmentation combined with the phases in time.The main improvements of this method are reflected in two aspects.Firstly,in view of the different amount of information provided by the different time periods of action,an improved time-phased video down sampling method based on Gaussian model is proposed.In this method,a smaller sampling interval is used to obtain more feature information while the action is in progress,and a larger sampling interval is used at the beginning and the end of the action to avoid data redundancy,which is similar to the Gaussian model,making the character of human action more distinct.Secondly,this paper proposes an interactive recognition framework that combines videos as a whole and individual segmentation.The overall video containing both sides of the action focuses on the extraction of global features such as relative position and orientation,and the individual segmentation video containing only a single person pays more attention to the description of the individual detailed action features.The framework uses the parallel multi-feature fusion network algorithm in the feature extraction phase.In the stage of action classification,the overall classification of the action and the classification results of the individual segmentation are combined at the decision level to improve the accuracy of human interaction action recognition.