摘要
提出一种单目视觉下在线识别目标三维行为的方法.该方法用匹配的标记点估计帧间相似变换,然后转换相似矩阵到对数空间以获取一致的四自由度运动参数序列.为解决持续时间敏感问题,提出基于多边形近似算法的时间尺度不变特征,并用动态规划实现特征序列的在线提取.在行为识别阶段,基于动态时间规整训练有限类别行为模板用于匹配测试行为序列.实验结果表明,该行为模板较对比方法类别可分性平均提高60%以上,并且可用于在线识别连续视频中的未知行为.
We present an approach to classify 3D behaviors online under monocular vision. We estimate similarity transformation between frames by matched markers, then transforms the similarity matrixes to logarithmic space to generate unified parameter sequence with 4 degrees of freedom. To eliminate the sensitivity of duration time, we formulate a time-scale invariant feature （TSIF） based on polygonal approximation algorithm, and implement online feature picking- up with dynamic programming. In the recognition phase, we use dynamic time warping to train the behavior templates with limited categories then recognize the test sequences. The experimental results show that the class separability of the proposed behavior template is increased by at least 60 % to the comparative approaches, furthermore, recognizing unknown behaviors in continuous video online is achieved.
出处
《自动化学报》
EI
CSCD
北大核心
2014年第8期1644-1653,共10页
Acta Automatica Sinica
关键词
三维重构
姿态估计
时间尺度不变特征
模板匹配
行为识别
3D reconstruction, posture estimation, time-scale invariant feature （TSIF）, template matching, behaviorrecognition
