【摘要】 现有时序动作定位算法大多依赖于时域建模，对频域特征的利用不足；此外，多尺度特征融合过程中也缺乏深层语义特征与浅层细节特征的双向交互，限制了算法的建模能力。对此，提出基于自适应谱滤波与双向门控聚合网络（adaptive spectral filtering and bidirectional gated aggregation network, ASFGNet）的时频双流动作定位算法。首先，设计自适应谱滤波模块，学习目标动作的频率成分，形成对时域特征的有效补充；其次，在时频交叉注意力的基础上提出双向门控聚合模块融合多尺度语义特征，提高算法对多尺度动作的表征能力。在THUMOS14和ActivityNet v1.3两个基准数据集上的对比分析表明，ASFGNet的平均mAP(mean average precision)指标分别达到68.3%与36.7%，优于多种主流算法；消融分析与收敛性分析验证了各模块的有效性与训练稳定性，定性分析进一步验证了算法的可解释性。更多还原

【Abstract】 Most existing temporal action localization algorithms rely on temporal modeling, which results in insufficient utilization of frequency domain features. Furthermore, during multi-scale feature fusion, there is a lack of bidirectional interaction between deep semantic features and shallow detail features, thereby limiting the algorithm′s modeling capability. To address these limitations, a timefrequency dual-stream action localization algorithm based on adaptive spectral filtering and bidirectional gated aggregation network(ASFGNet) was proposed. Firstly, an adaptive spectral filtering module was designed to learn the frequency components of the target actions, effectively supplementing the temporal features. Secondly, based on cross-attention mechanisms in the time-frequency domain, a bidirectional gated aggregation module was introduced to fuse multi-scale semantic features, enhancing the algorithm′s ability to represent actions across varying scales. Comparative analysis on the THUMOS14 and ActivityNet v1. 3 benchmark datasets demonstrates that ASFGNet achieves average mean average precision(mAP) scores of 68. 3% and 36. 7%, respectively, surpassing several mainstream algorithms. Ablation studies and convergence analysis validate the effectiveness of each module and training stability,while qualitative analysis further verifies the interpretability of the algorithm.更多还原