融合多姿势估计特征的动作识别

目的 为了提高静态图像在遮挡等复杂情况下的动作识别效果和鲁棒性,提出融合多种姿势估计得到的特征信息进行动作识别的方法。方法 利用已得到的多个动作模型对任意一幅图像进行姿势估计,得到图像的多组姿势特征信息,每组特征信息包括关键点信息和姿势评分。将训练集中各个动作下所有图像的区分性关键点提取出来,并计算每一幅图像中区分性关键点之间的相对距离,一个动作所有图像的特征信息共同构成该动作的模板信息。测试图像在多个动作模型下进行姿势估计,得到多组姿势特征,从每组姿势特征中提取与对应模板一致的特征信息,将提取的多组姿势特征信息分别与对应的模板进行匹配,并通过姿势评分对匹配值优化,根据最终匹配值进行动作分类。结果 在两个数据集上,本文方法与5种比较流行的动作识别方法进行比较,获得了较好的平均准确率,在数据集PASCAL VOC 2011-val上较其他一些最新的经典方法平均准确率至少提高近2%。在数据集Stanford 40 actions上,较其他一些最新的经典方法平均准确率至少提高近6%。结论 本文方法融合了多个姿势特征,并且能够获取关键部位的遮挡信息,所以能较好应对遮挡等复杂环境情况,具有较高的平均识别准确率。

Fusing multiple pose estimations for still image action recognition

Objective To adapt to occlusion or other complex situations, an action recognition method is proposed which fuses multiple pose estimation features.Method Multiple pose features will be obtained using multiple action models. Each pose feature information includes key point positions and pose scores. Distinguishing key pointsare extracted from all train images and computing relative distances between point pairs. An action template is built using all features of the train images of the action. Multiple feature information consistent with multiple action templates are extracted from each test image from multiple pose features.Multiple features of the test image are-matched with the corresponding action template and then matched values are optimized using pose scores. Result The experimental results have shown that the average accuracy of the proposed method is approximately 2% better than some other state-of-the-art methods on VOC 2011-val set, and is approximately 6% better than some other state-of-the-art methods on Stanford 40 actions set. Conclusion By fusing multiple pose features, the proposed method can adapt to occlusion and other complex situations and improve average recognition accuracy.