This paper introduces an enhanced behavioral recognition algorithm,YOWO-Uni,which inherits the framework of the YOWO algorithm but reconfigures its components for improved performance.Firstly,the 3D network branch's 3D-ResNext-101 is substituted with UniFormer-XS,thereby augmenting the algorithm's capability to extract temporal information.Secondly,the LSKA attention mechanism is incorporated into the 2D network branch,enhancing the extraction of spatial features.Further-more,lightweight Ghost convolutions are employed to reconstruct the channel fusion attention module,effectively reducing redun-dancy and lowering the model's parameter count.Lastly,the EIOU loss function is adopted to bolster the stability of bounding box regression.Experimental results on the UCF101-24 and J-HMDB-21 datasets affirm that YOWO-Uni significantly mitigates model complexity while concurrently elevating its representational capacity.
关键词
YOWO/人体行为识别/注意力机制/Ghost/损失函数
Key words
YOWO/human action recognition/attention mechanism/Ghost convolution/loss function
万方数据