Action recognition of body pose from static image is exploited in this paper. A hierarchical part tree structure is proposed. In the structure, each node is composed by a collection of poselets to represent its pose variation and pairs of linked nodes are constrained to form a pictorial structure. Grounded on the structure, a discriminatively trained action recognition model based on hierarchical part tree is presented. Except for deforming cost, the pairwise potential function in the model introduces co-occurrence cost. Parent part contains child part and the relative position of linked nodes is described by normal distribution, and thus the matching procedure is inferred efficiently in the framework of distance transform and message passing. Three models with different number of nodes by trimming the tree are comparatively evaluated on two datasets. Experimental results demonstrate that coarse parts in former two layers have strong saliency for action recognition, the recognition capability is further improved by body parts in the third layer, and the anatomical stick parts in the fourth layer are basically not useful for action recognition.%研究从静止图像中识别人体姿态动作.首先提出层次部件树结构,树中每个节点由一组Poselet表示该肢体部件的姿态变化,节点之间相互制约,构成一个Pictorial结构.基于此结构,提出基于层次部件树结构的动作识别判决模型.Pictorial结构的对偶潜在函数中除了变形代价,引入Poselet同时出现代价.由于树的邻接节点之间存在包含关系,相对位置可以使用高斯分布描述,推理过程沿用距离转换和置信度传播算法,实现高效匹配.在2个数据集上,对剪枝后节点数量不同的3种判决模型的实验表明,前两层的粗粒度节点具有较强的动作识别显著性,第三层进一步提高动作识别能力,第四层的原子部件对动作识别无明显作用.
展开▼
