تشخیص عمل براساس شبکه بیزی دینامیکی سلسله مراتبی Action recognition based on hierarchical dynamic Bayesian network

1 Introduction In recent years, human action recognition has become a core issue in field of computer vision. Because of complexity and uncertainly, action identification is still a very challenging subject. Many action recognition methods tend to design descriptors and classifying based feature matching [11, 13]. The previous action recognition methods main include two classes, i.e. feature description and action classification. According to [28], feature representation is always key task for recognizing actions. In general, the feature presentation usually is divided into global representations and local representations. The global feature records total image presentation, however, the global feature is often disturbed by occlusions, viewpoints changing and noises. The global-based feature includes optical flow-based presentation [27], silhouette-based descriptor [3], edge-based features [34], and motion history image (MHI) [2], and so on. The local feature always describes patches independently, and the patches are combined together to build space–time model [17], such as HOG [4] and SURF [1]. The local descriptor presents action video more effectively, especially for noises images and partial occlusions images. However, processing related interest points is high time cost. In this paper, we present a hierarchical graph model framework for meeting complex videos semantic identification. The highlights of this method include two points, one is hierarchical action semantic dictionary construction, and another is hierarchical dynamic Bayesian network semantic inference model. The motive of the proposed approach is to recognize actions with the higher accuracy. We consider the problem based 3 points: (1) for reducing video dimension and recognition time cost, we use deep neural networks [19] to extract video features firstly, and the Aligned Cluster Analysis (ACA) [35] is utilized to get representation frames. As we known, the selected discriminating features always are better performance. Based the deep neural network and the ACA, the better discriminating features are got from original action video. (2) For enhancing robustness of recognition, we propose to construct the hierarchical action semantic dictionary (HASD). As we known, high level semantic analysis is always very important for complex and uncertain identification problem, dictionary-based classifier has better performance to recognize action. In this paper, we propose dictionary-based recognition to enhance recognition robustness. (3) At the same time, it is proved that probability graph model is an efficient tool to dig hidden state information [29]. The dynamic Bayesian network (DBN) is promising method to present random time series signals entirely. Hence, we select the HDBN to present action. The HDBN-based signals processing can accomplish 2 tasks: one is to present action entirely and clearly, two is to dig more hidden semantic state information. (4) Based on the above-mentioned, we combine high level semantic analysis and graph model together to obtain effective representation. Based on the HDBN inference and semantic analysis, and the HDBN + HASD-based method can finish recognition task effectively. The rest of this paper is organized as follows. The related works is described in Section 2, and the proposed approach is described in Section 3. Section 4 compares proposed method with the existing models. Finally, conclusions are given in Section 5.