Research And Application Of Giant Panda Action Reconition

Action recognition is an important direction in the field of computer vision,it is also a highlight in recent years with the prosperity of deep learning.Based on the image and video information,the action recognition technology extracts the spatial and temporal characteristics for modeling,so as to identify the actions of target.It brings convenience to decision makers to respond quickly while take corresponding measures.Current action recognition techniques and related researches mainly focus on human actions.The definition of human action is more accurate and recognizable,and the data set is sufficient.In recent years,human action recognition technology has made a gratifying progress.Compared with researches on human action detection,there exist following urgent problems in giant panda action recognition:(1)Data on giant panda action is particularly sparse.Due to the effect of panda habits,there exist high similarity in different giant panda videos.(2)There is no standard dataset to evaluate the performance of panda action recognition.(3)Because of factors like pandas' body posture,the visual information of giant panda is ambiguous,which increases the difficulty in action recognition.(4)In terms of long-term development,giant panda behavior recognition technology has certain requirements on time performance.To solve above problems,this thesis proposes two action recognition algorithms for giant panda based on Transformer model.The first one is called CNN-transformer algorithm.Based on the space-time information model,it utilizes the convolutional neural network to extract the spatial feature map of video frames,then compress the feature map into a vector and input it into the encoder part of the Transformer to realize the serialized modeling.The other one is called ConvTransformer algorithm,which integrates convolution operation in Transformer model.This algorithm innovatively improves the calculation process of Transformer and directly receive feature graph as input,thus making advantage of spatial information.In order to evaluate the algorithm as accurately as possible,we collect some panda videos to construct a panda short video dataset with five action types.Meanwhile,in order to impartially evaluate proposed algorithms,a public human action dataset UCF101 is utilized in our comparison schemes.Extensive experimental results demonstrate that the proposed algorithms perform well in both the UCF101 dataset and the panda short video dataset.Besides,this thesis implements a video retrieval platform based on the giant panda action recognition algorithm.This platform can automatically implement video classification according to users' upload data,then retrieve top 50 most similar videos and return to corresponding users.