基于多示例学习法的人体行为识别.pdfVIP

2016~剁翩 ú7 信息技术 Jφ岛，回曲，男必叫，.， 文章编号:1α)9 -2552(2016)07 -0065 -06 001:10. 13274/j. c咏i. hdzj.2016.07.018 基于多示例学习法的人体行为识别 王军1 ， 2 (1.电子科技大学中山学院，广东中山 528402; 2. 中南大学信息科学与工程学院，长沙41∞175) 摘 要:提出了基于多示倒学习法的人体行为识别方法。利用人体骨架模型，将人体主要关节 的属性特征作为人体运动的几何特征，提出了一种基于行为几何特征的自适应行为分解算法， 将行为分解为简单动作。把分解后的行为看作一个包，各个动作看作包中的各个示例，结合多 示例学习法与AnyBoost 算法提出了多示例行为学习算法(MILBoost 算法) ，通过在多示例框架下 对每一类行为进行学习，得到强分类器用于未知行为包的识别。实验结果表明该方法与其他方 法相比具有更高的识别精度，并且在有噪声或干扰的情况下具有很好的识别精度。 关键词:行为识别;多示例学习;行为分解;包 中图分类号: 1凹91 文献标识码:A Human activity recognition using multiple instance learning WANG Junl ,2 (1. Zho耶han Ios侃侃te ， University of Elec剧，nic Science and Tecbno问野， Zho吨抽an 528402 ,Guangd幅g pro世阻，由ÎDa; 2. Scbω，1 of Ioformation Science and Engineering , Central South University , chan伊ha41∞75 ，Cb旭a) Abstract : A novel method hased multiple instance learning is proposed for human activity reco伊tion. Based on the skeleton model of human , a sequence major joints features 缸e used 跑出e geometric characteristics of human motion and a adaptive segmentation algorithm is introduced to de∞mpo回 activity into simple actions. The decomposed activity comprises laheled hags 出at 缸e composed of unlaheled instances comprising to action. MILBoost algorithm is presented hy comhine multiple instance learning literature 明白 the AnyBoost fr部nework. Laheled hehaviors are used to 位am a s田ng classifier which is used ωpredict the lahels of unseen hehavior hags. 四le experimental results show the effectiveness of the proposed method in comparison 响曲 other related works in the literature and can also tolerate noise and interlerence conditions.