Assigning PLS Based Descriptors by SVM in Action Recognition

摘要

In this paper, we propose assigning PLS based descriptors by SVM to obtain the representations of human action videos. First, in addition to the spatially gradient orientation, we add spatio-temporal gradient statistic to generate the extended Histogram of Oriented Gradient (HOG). Second, different from requently-used cuboid descriptors in which Principal Component Analysis (PCA) is applied for dimension reduction, the proposed features utilize the Partial Least Squares (PLS) method for better performance. Then, we apply a multi-class SVM for assignment instead of assigning descriptors to the nearest (Euclidean distance) visual word in traditional Bag of Visual Words (BOVW) framework. Finally, the K-nearest neighbor algorithm is used to classify the histogram of visual words. The experimental results on the facial expression dataset and KTH human activity dataset validate the effectiveness of our proposed method.