基于累积边缘图像的现实人体动作识别

为了从现实环境下识别出人体动作,本文研究了从无约束视频中提取特征表征人体动作的问题. 首先,在无约束的视频上使用形态学梯度操作消除部分背景,获得人体的轮廓形状; 其次,提取某一段视频上每一帧形状的边缘特征,累积到一幅图像中,称之为累积边缘图像 (Accumulative edge image, AEI); 然后,在该累积边缘图像上计算基于网格的方向梯度直方图(Histograms of orientation gradients, HOG),形成特征向量表征人体的动作, 送入分类器进行分类. YouTube数据集上的实验结果表明,本文的方法比其他方法更加有效.

Recognizing Realistic Human Actions Using Accumulative Edge Image

The problem of extracting feature from unconstrained videos for representing human actions has been investigated in order to recognize human actions in complex environment in this paper. Firstly, morphological gradient was used to eliminate most background information. Then, edge of shape was extracted and accumulated to a frame, which was named accumulative edge image (AEI). Grid-based histograms of orientation gradients (HOG) were calculated and formed a feature vector that captured the characteristic of human actions in this video sequence. Using support vector machine (SVM), the method was tested on the YouTube action dataset. The obtained impressive results showed that this method was more effective than other methods in YouTube action dataset.