Animal Action Recognition Based On Cross-correlation Fusion And Discriminative Filters

Animal behavior analysis is an important research direction of the intersection of biology and computer science.It mainly makes use of computer vision and artificial intelligence technology to automatically analyze the behavior of animal body,joints and other parts.As the focus of animal behavior analysis,animal video behavior recognition needs to extract the key visual features of the action sequence in video,and then identify the spatio-temporal information of the target action through machine learning,deep learning and other algorithms,so as to complete the visual content understanding and pattern recognition of the target action.This paper focuses on the understanding and analysis of spatio-temporal action features in video information,and mainly studies the video behavior recognition algorithm involved in it.In order to solve the problem that the traditional two-stream network can't learn the non-local spatio-temporal dependence and the single-stream network can't recognize the key region,the cross-correlation fusion network and the discriminant filter network are proposed respectively.The main research and work of this paper are as follows:1.Aiming at the defect that the traditional two-stream network can't perform non-local feature fusion,action recognition algorithm based on cross-correlation two-stream fusion network is proposed.Two-stream network obtain the spatial and temporal features respectively by learning the image static information and optical flow motion information,and use the late-fusion method to combine the both outputs for improving recognition performance.However,this fusion approach can't learn the non-local dependencies of these two features,e.g.relationship of features of "object" and "action" in different regions.The cross-correlation network proposed in this paper computes the global response between the two features by introducing the cross-correlation matrix,so as to capture the non-local spatio-temporal dependence.Finally,it achieved 95.0% accuracy in the UCF101 human dataset and 99.1% accuracy in the rat behavior dataset.2.Aiming at the defect that the global average pooling layer cannot effectively capture the discriminant region in the fine-grained behavior,action recognition algorithm based on discriminant filter single-stream network is proposed.Deep neural network extracts visual features by convolution layer and classifies them by global average pooling and full connection layer.However,this method cannot consider the discriminant region in the target features.In this paper,global maximum pooling layer and adaptive attention module were used to detect the discriminating region in visual features,so as to enhance the generalization of the network.Finally,98.1% accuracy was achieved in the rat behavior data set and 79.4% accuracy in the larva behavior data set.