摘要
人体姿态识别是当前自动视频理解技术的研究热点,其难点在于在实际应用中很难同时保障准确度、鲁棒性和实时性.当前基于二维图像的主流算法中,一类为基于高层人体结构的信息,其准确度高,但实时性较差;另一类为基于低层图像信息,算法简单,但其准确度较低.针对该问题,文中提出一种人体姿态建模和识别算法.该算法首先采用高斯混合模型快速提取运动目标和归一化轮廓图像,然后利用人体轮廓参数构建一组12维特征向量,建立人体姿态模型,最后通过分层识别方法实现对人体姿态的认知.该算法可以有效地识别人体姿态,计算复杂度较低,对存在干扰的图像具有较好的识别效果.基于标准视频库的实验结果验证了方法的有效性,与链码标记算法的对比实验验证了方法的优越性.
Human posture recognition is a research hotspot in the automatic video understating technology. It is difficulty to ensure the accuracy, robustness and real-time at the same time in practical applications. The existing mainstream algorithms based on 2D image information can be classified into two classes： the methods based on high level human model which have high accuracy and high-complexity; the methods based on low level image information which have low-complexity and low accuracy. A algorithm for human posture recognition is proposed to solve this problem. Firstly, Gaussian mixture model is exploited to extract foreground and normalized human silhouette. Then, a 12-dimensional invariant eigenvector is constructed, thereby the human posture model is established. Finally, a hierarchical recognition method is adopted to recognize the postures. This algorithm is efficient, has low complexity, and achieves good effect for some interfered images. The results of the experiments based on standard video database verify the validity of the proposed algorithm, and the superiority of the algorithm is also verified compared with chain code algorithm.
出处
《模式识别与人工智能》
EI
CSCD
北大核心
2011年第3期411-416,共6页
Pattern Recognition and Artificial Intelligence
基金
国家863高技术发展计划项目(No.2009AA11A113)
高等学校博士学科点专项科研基金项目(No.20093402110014)资助
关键词
人体姿态识别
人体姿态模型
码书
傅立叶描述子
Human Posture Recognition, Human Posture Model, Codebook, Fourier Descriptor
