人体行为识别的Markov随机游走半监督学习方法

Human Action Recognition Using Markov Random Walk Based Semi-supervised Learning

针对目前人体行为识别方法大都需要大量有标注样本的问题,提出一种基于Markov随机游走的半监督人体行为识别算法.首先提取序列图像各帧人体区域的网格统计特征,再采用基于对手惩罚策略的竞争神经网络对其进行聚类和编码,将图像序列表示的人体行为变换为符号序列;然后根据行为之间的归一化编辑距离建立已标注行为、未标注行为和类别之间的Markov链,并采用Markov随机游走过程来预测未标注行为的类别;最后采用最大后验概率准则对观测到的未知行为进行分类.对Weizmann数据集中人体行为的识别实验结果表明,该方法是一种有效的人体行为识别方法,在标注样本很少的情况下平均识别精度可以超过80%.

To accommodate the great challenge of obtaining accurate and detailed annotations of training data in human action recognition, in this paper we propose a semi-supervised algorithm based on Markov random walk. We first calculate mesh features of each image in the human action sequence. Then the mesh features are vector quantized through a rival penalized competitive neural network and the actions described by time-sequential images are converted into symbolic sequences. We then construct a Markov chain to reveal the labeled actions, unlabeled actions and action categories in training set according to their edit distance. Furthermore, we predict class label of unlabeled actions through a procedure of Markov random walk. The test sequence is then classified by the maximum posteriori probability criteria. The experiments on Weizmann dataset demonstrate the effectiveness Of our method. The average recognition accuracy can exceed 80% even when only very small amount of labeled actions acquired.