基于稠密光流轨迹和稀疏编码算法的行为识别方法

针对现有行为特征提取方法识别率低的问题,提出了一种融合稠密光流轨迹和稀疏编码框架的无监督行为特征提取方法(DOF-SC)。首先,在稠密光流(DOF)轨迹提取的基础上,对以轨迹为中心的原始图像块进行采样作为轨迹的原始特征;其次,对轨迹原始特征基于稀疏编码框架训练稀疏字典,得到轨迹的稀疏特征表示,利用词袋(BF)模型对稀疏特征聚类得到轨迹的码书,再根据码书对每个动作中出现的所有轨迹所属的码书类别进行投票,统计该动作中每个码书出现的次数,得到行为特征;最后,对行为特征利用基于直方图交叉核函数的支持向量机(SVM)进行训练得到行为识别模型,再利用该模型对行为进行分类预测,得到最终行为识别的结果。在对轨迹采样10%的情况下,DOF-SC算法得到的行为识别准确率在KTH数据库上高出采用运动边界直方图(MBH)作为特征的行为识别准确率的0.9%,在YouTube数据库上高出MBH作为特征的行为识别准确率的1.2%。实验数据表明了所提方法对行为识别的有效性。

Action recognition method based on dense optical flow trajectory and sparse coding algorithm

Focusing on the issue that the existing action feature extraction method achieves lower recognition rate, a novel unsupervised one for action recognition by combining Dense Optical Flow trajectory and Sparse Coding (DOF-SC) algorithm was proposed. First of all, the trajectory-centered image patches were sampled as the original features based on the extraction of Dense Optical Flow (DOF). Then, the sparse dictionary on the basis of Sparse Coding (SC) framework was trained, and the sparse feature representation of the trajectory through dictionary was got, then the code book of the trajectory by clustering with the Bag-of-Feature (BF) model was achieved, and the trajectory of every action by code book was voted, the action features by counting the number of every code book were got. Finally, the examples for action recognition was classified and predicted by Support Vector Machine (SVM) with the kernel of histogram intersection function. The accuracy of the DOF-SC algorithm is superior to the accuracy of Motion Boundary Histogram (MBH) as the action feature by 0.9% in the KTH (Kungliga Tekniska Hgskolan) database and 1.2% in the YouTube database with the trajectory sampling rate of 10%. The results prove the effectiveness of the unsupervised action feature extraction method.