Improved Slow-Fast Network For Panda Action Recognition

In the field of computer vision today,as one of the most popular subjects,action recognition is widely concerned by researchers.Compared to human actions,works done about panda action recognition tend to be insufficient.Panda is a unique animal species in China.Panda action recognition would promote the development of relative subjects and breeding of pandas.In this thesis,the problem of animal behavior detection is studied based on deep learning methods.Our team collected panda behavior videos in the early stage,and established a panda behavior dataset(Panda Behavior Analysis,PBA).It includes bounding boxes targeting pandas and 4-layer behavior label groups(each group contains9,9,8,and 5 types of behaviors labels,respectively).We applied Yolo v5 to locate the bounding boxes of panda on the original videos,and then manually adjust the bound boxes.This method accelerated the establishment of dataset and earned considerable time for later work.Yolo v5 network is used to detect the attitude layer of giant pandas.The detection results basically reach the expected value beforehand.At the same time,for the behavior labels with time dimension,this thesis also proposes a behavior detection method based on convolution vision transformer(Cv T)improved Slow-Fast network to identify the behavior of giant pandas,so as to help staff quickly identify abnormal behaviors and reduce labor costs,and can also assist researchers to analyze the behavior of giant pandas to ensure the stability of giant pandas during the breeding period.The network utilizes the characteristics of the fast and slow channels in the Slow-Fast network,divides the input panda video image frames into two channels,fast(16FPS)and slow(2FPS),and extracts features through the 3D convolution layer in the two channels respectively.The multi-head self-attention mechanism is then applied to calculate the panda behavior classification probability in the input video image frame.The mean average precision(m AP)obtained by this method reached 46.4%.Slow-Fast proves the similar result when tested on Kinetics-400.This proves the Cv T Slow-Fast is capable on panda action recognition.