The Research On Athlete Behavior Recognition Based On Deep Learning

Human behavior recognition is one of the fundamental research directions in computer vision,which is concerned with recognizing human behavior in videos and classifying them.Sports videos contain rich information,and indexing sports videos by sports category is an important tool for post-processing such as post-game analysis and coaching tactics formation,which has great commercial potential and application value.In order to improve the performance of athlete video classification,the main research of this paper is as follows:(1)To address the problem that the cross-fusion of spatial and temporal features at multiple scales cannot be fully considered and the correlation information of temporal and spatial features is lost,this paper proposes the Multiscale Spatio-temporal Pyramid(Multiscale Spatio-temporal Pyramid,MSP),which adds deformable convolution to the feature pyramid to capture multi-scale features and effectively improve the modeling ability of target deformation,and adopts the best fusion strategy weighted fusion to fuse the obtained spatio-temporal features,so as to construct a more efficient behavior recognition model.In addition,for the problem of high model complexity and large number of parameters in 3D convolution-based behavior recognition methods,a lightweight 3D Dense Net module(3D Dense Net module,DDense Net3D)is proposed,introducing dynamic grouping convolution to improve the dense blocks in its dense network.It is able to facilitate the reduction of the number of model parameters without destroying the original network structure.The experimental results show that MSP has significantly improved the accuracy of athlete behavior recognition on Sports8 dataset and Olympic16 dataset;DDense Net3 D has greatly reduced the number of parameters and computation while ensuring the accuracy of athlete behavior recognition on Sports8 dataset and Olympic16 dataset.(2)To address the problem that the lightweight network model can reduce its expressive power and the interference caused by giving the same weights to different regions and channels,this paper proposes an improved attention mechanism for athlete behavior recognition.Since the serial connection of spatial attention module and channel attention module in CBAM attention mechanism will generate interference,the attention mechanism of Parallel Addition Convolutional Block Attention Module(Parallel Addition Convolutional Block Attention Module,PCBAM)is proposed,and in order to further improve the athlete behavior In order to further improve the accuracy of athlete behavior recognition,PCBAM is added between and within the dense blocks of DDense Net3 D to optimize DDense Net3 D,and two new models PCBAM-DDense Net3D-In and PCBAM-DDense Net3D-On are proposed.DDense Net3D-On achieved 96.3% and 79.8% accuracy of athlete behavior recognition on the Sports8 dataset and Olympic16 dataset,respectively.Compared with network models such as C3 D,Res Net3 D,and Two-Stream,it has faster convergence,better classification performance,and certain generalization.