基于流密度的多重交互集体行为识别算法

发布时间：2018-04-22 02:06

本文选题：集体行为 + 多重交互　；参考：《计算机学报》2017年11期

【摘要】：作为当前模式识别领域的研究热点之一,识别视频场景中的集体行为有着重要的科研意义和应用价值.已有的识别方法都忽略了对集体行为中多重交互关系的描述和分析,使得复杂视频场景中的集体行为识别仍存在局限性.该文提出了一种基于流密度的多重交互集体行为识别算法FDA,能够识别集体行为的局部和全局模式.为了准确地衡量个体间的行为一致性程度,受流体力学启发定义了流密度,并通过基于流密度的聚类算法识别局部子群组;为了精确描述集体行为客观存在中的多重交互关系,提出了多重邻接关系模型.该模型能够分析子群组间的全局一致性,进而通过合并具有较高一致性的子群组得到全局集体行为模式.在多个真实视频数据集上的实验结果表明,FDA算法相比于已有方法具有更高的识别精度.
[Abstract]:As one of the research hotspots in the field of pattern recognition, the recognition of collective behavior in video scene has important scientific significance and application value. The existing recognition methods ignore the description and analysis of multiple interactions in collective behavior, which makes collective behavior recognition in complex video scenes still limited. In this paper, a multi-layer interactive collective behavior recognition algorithm based on flow density, FDA-based, is proposed, which can recognize the local and global patterns of collective behavior. In order to accurately measure the degree of behavior consistency among individuals, the flow density is defined by fluid mechanics, and local subgroups are identified by clustering algorithm based on flow density, and in order to accurately describe the multiple interactions in the objective existence of collective behavior. A multiplex adjacency model is proposed. The model can analyze the global consistency among subgroups, and then obtain the global collective behavior pattern by merging subgroups with high consistency. The experimental results on many real video data sets show that the FDA-based algorithm has higher recognition accuracy than the existing methods.
【作者单位】：郑州大学信息工程学院;
【基金】：国家自然科学基金(61170223,61502434,61672469,61772475) 河南省科技攻关项目(172102210011)资助~~
【分类号】：TP391.41

【相似文献】