基于关键帧的复杂人体行为识别

摘要

A key frames based method is proposed for recognition of complex human behavior. The body contours are adopted to represent actions. The complex behavior fragment is divided into simple behaviors by the shot boundary detection algorithm. To improve the recall rate, double sampling is employed during the segmentation. The self-splitting competitive learning algorithm is utilized to acquire key frames of simple behaviors. Finally, the recognition of complex human behavior is achieved according to the similarity of the key frames. During the process, the visual factor, the order factor and the interference factor are taken into account. Thus, the calculations of the recognition are more rational and comprehensive. The proposed method is verified on UCF Sports database and self-built database and the results indicate the high recognition accuracy of the proposed method.%提出基于关键帧的复杂人体行为识别算法。采用人体轮廓表示人体动作，根据轮廓特征向量采用镜头边界检测将复杂行为片段分解后得到一系列简单行为，分解过程中采用二次采样提高分界点的查全率，采用自分裂竞争学习提取简单行为片段的关键帧。最后根据行为片段中关键帧的相似度进行复杂人体行为的识别，识别过程中综合考虑视觉因子、顺序因子及干扰因子，使行为识别的计算更合理全面。在UCF Sports数据库及自建数据库上的实验表明文中算法具有较高的识别精度。