基于多时间划分的深度聚合特征的行为识别

Action recognition based on deep aggregation feature from multi-temporal segmentations

重点研究了视频中的人体行为识别问题,提出了基于多时域划分的深度聚合特征的行为识别算法。通过多时间划分的采样形式,丰富行为深度特征的多样性,并引入多聚合中心,量化深度特征与聚合中心的残差关系,进而形成人体行为的结构化建模方法,构建可以学习复杂行为特征分布的深度聚合模型,解决时间跨度较长时的人体行为分类问题。通过在开源行为数据集UCF101与HMDB51上的实验,验证了算法的有效性和先进性。

This paper focused on the problem of human action recognition in videos,and proposed an algorithm based on deep aggregation feature from multi-temporal segmentations.It adopted the multi-temporal segment based sampling method to enrich the diversity of deep features of actions,furthermore employed the multi cluster centers and quantified the residual relations between deep features and cluster centers to build the deep aggregation model which was capable of learning the feature distribution of complex human actions by researching on the structured modeling methods of action.The purpose of this paper was to address the problem of human action recognition with a long-range span.It devised the experiments on the UCF101 and HMDB51 datasets to verify the effectiveness and advancement of this algorithm.