Research On Behavior Recognition Algorithm Based On 3D Convolutional Neural Network

Multimedia on Internet is growing rapidly,resulting in an explosion in the number of videos being shared.The challenge we all face how to sort through all these videos,figure out what they're about,and enable people to find the ones they are interested in.The computer vision community has worked on video analysis for decades and tackled different problems.Currently,we lack a generic feature descriptor for videos,which can be used for a variety of video processing problems.Such a generic feature descriptor helps in solving video analysis tasks in a homogeneous way,thus also enables large-scale video applications.Deep learning is a research hotspot in recent years.In deep learning,the convolutional neural network is a typical neural network,it has superiority to the traditional neural network in the recognition effect,and it does not require manual design features.This article analyzes and studies the representative C3 D neural network in the action recognition,discusses and improves the existing deficiencies,and proposes a spatio-temporal filter.The main work and innovation are as follows:1.Constructing a spatio-temporal cascaded convolution module---Fake-3D module.Inspired by the theory of tensor CP decomposition,we constructed a linear superposition of adjacent features on the spatial-temporal domain by 3×1×1 convolution,1×3×1 convolution,and 1×1×3 convolution.Concatenating spatio-temporal convolutions to simulate a 3×3×3 convolution operation on the original spatio-temporal domain.We name this spatio-temporal concatenation convolution as Fake-3D.2.Proposing a new convolutional neural network architecture,namely Fake-3D neural network.Replacing the 3D convolution filter in the original C3 D neural network with our constructed Fake-3D module.3.Exploring Fake-3D blocks of different structures and finding a best performing Fake-3D block.The idea of the Fake-3D network is to learn the spatiotemporal characteristics through feature fusion of multiple low-rank cascade convolution kernels,that is,in the Fake-3D block,each subfilter uses the extracted features and the original features of the same module as input.Finally,experiments on two benchmark datasets demonstrate the effectiveness of our new architecture.