Research On Spatio-temporal Graph Convolution Behavior Recognition Algorithm For Intelligent Surveillance Scene

As a key problem in the field of computer vision,accurate recognition of the semantic information embedded in behaviour can facilitate people’s daily lives and has a very important impact on all aspects of society.Existing behaviour recognition methods can be classified into image sequence and human skeletal sequence methods according to the type of data.In this paper,a graphical convolutional behaviour recognition method based on spatio-temporal features is proposed for surveillance video data.Firstly,this paper constructs a spatio-temporal graph convolution model based on the framework of the classical spatio-temporal convolution model,which is able to extract the spatio-temporal features of the input video sequences and construct the graphs.The problem of inadequate extraction of spatio-temporal features in the classical spatio-temporal convolution model is solved,and the expressiveness and extraction ability of the graph structure for spatio-temporal features is verified by experiments.Secondly,to solve the problem of multi-target action recognition in image sequences,this paper constructs a multi-target action recognition model based on fast slow convolution and graph convolution network.This model can detect the actions of multiple targets to recognize targets at the same time.It solves the problem that the classical behavior recognition method for video data ignores the relationship between multiple objects.Finally,the algorithm in this paper has been tested on UCF101,HMDB51,kinetics400 and AVA datasets respectively,and has good performance in action recognition and spatio-temporal action detection tasks.The algorithm is applied to dangerous behavior recognition in the maintenance environment of railway locomotive depot through self-made datasets.