Aiming at the problem that existing spatio-temporal action detection algorithms decouple the localization and classification tasks into a single stage

resulting that the model cannot be optimized uniformly

a spatio-temporal action detection algorithm based on dual-branch feature fusion

YOWO-cross-channel attention aggregation(CCAA)

is proposed.The model adopts a unified architecture of dual-branch parallelism

3D convolutional neural network(CNN)branch is used to extract the spatio-temporal features of the action

and 2D CNN branch is used to extract the spatial features of the motion body.The feature fusion module CCAA is designed to realize the dual-branch cross-channel information interaction

and global dependencies in spatial and channel dimensions are respectively captured through the self-attention mechanism to highlight the action subject.Finally

SIoU is adopted as the bounding-box regression loss function to accelerate the model convergence speed.Experimental results on public datasets show that the improved algorithm achieves 86.81% of F-mAP and 51.43% of V-mAP on UCF101—24

and 74.17% F-mAP and 85.7% V-mAP on JHMDB-21 dataset

respectively.In addition

the frame rate reaches 26.5fps

which achieves better results in accuracy and real-time property.