基于线性注意和类别关联特征学习的在线动作检测OA

To fully and reasonably utilize the contextual features of actions,category-related features and predicted future features for rapid action detection in online action detection,the online action detection method based on linear attention and category association feature learning was proposed.The Transformer architecture was improved by employing lightweight linear self-attention based on the Hadamard product to reduce computational cost for video contextual feature learning.The action features from training samples were clustered to associate the video sequence context with action category features for achieving the effective learning of category-associated feature representations.By integrating contextual features,category-associated features and predicted future features,the action discrimination at corresponding moments was enhanced.The performance experiments were conducted on typical datasets to realize the hyperparameter selection analysis,and the working accuracy and operational efficiency of different methods were compared.The ablation experiments and visualization analysis were provided.The results show that on the Thumos14(TSN-Anet),Thumos14(TSN-Kinetics)and HDD datasets,the mAP values of the proposed method are respectively improved by 0.2,0.5 and 0.2 percentage point compared with the Colar method,which indicates that the new method outperforms the currently advanced Colar method.