基于阶层多观测模型的多人行为识别

Hierarchical multi-observation model for multi-person activity recognition

为解决多人行为识别中高维特征空间、角色分配不准确和复杂的时间结构等问题,该文分析了多人行为的特点,提出了一种递归的多层随机网络模型。该模型通过多层网络表达行为的多尺度特性,并由高层体现行为的长时间依赖性。通过对观测的分解大大降低了特征空间的维数,从而降低了问题的复杂度,并在一定程度上消除了目标角色分配不准确带来的影响。实验结果表明:该文提出的模型比其他常用模型具有更好的识别效果,即使对复杂行为依然具有91.3%以上的识别率。

A reeursive multi-level stochastic model is presented for multi-person activity recognition, such as in a high-dimension feature space, role assignments and with complex temporal structures. The model represents the multi-scale characteristics of activities by the multi-level network and captures the long-term dependency using high-level chains. The model decomposing observations considerably reduce the dimensionality of the feature space. The assumption that sub-observations have uniform distributions partly eliminates the effects of role assignment errors. Test results demonstrate that the model outperforms other popular models, with 91.3% recognition rate even for complex activities.