人体行为识别的条件随机场方法

The Conditional Random Fields Method for Human Action Recognition

针对人体行为的时变性,提取人体运动的侧影序列作为描述行为的特征。利用条件随机场方法建立人体行为模型,并通过序列数据的标记解决行为识别问题。该方法特征提取简单,针对运动状态序列而非单帧图像进行建模,提高了识别准确率;同时对数据没有条件独立性假设,具有更加广泛的适用性。在视频行为数据库KTH上的测试结果表明:条件随机场优于隐马尔可夫模型和支持向量机,相对于已有方法更加简单易用,且识别准确率高于其他方法。

In this paper,we first extract the silhouette sequence of a moving person and describe its action by using the features extracted from its silhouette sequence based on the time-varying property of human action.Then we mathematically define the human action by using the conditional random fields and recognize different actions by labeling the video sequences.This method has the following advantages： it is easy to extract features from silhouette sequence and thus it could be used in human action recognition application on large-scale video sequences;the conditional random fields models the states of action nor separate frame images and thus it improves the classification performance;it avoids the conditional independence hypothesis and thus could be used on wide range of datasets.The experimental results on popular video based human action dataset,i.e.,KTH,show that the conditional random fields method outperforms both hidden Markov model and support vector machine.In addition,our method outperforms existing human action recognition methods.Although our method performs comparably to the state-of-the-art method,it is easier to use in practice.