This thesis explores the popular bag-of- features framework for human action recogni- tion. Different feature detectors and descrip- tors are described and compared.
The combination of the Harris3D detector and the HOG/HOF descriptor is chosen for use in the experimental setup. Unsupervised and su- pervised classification methods are compared to show the difference in performance. The supervised algorithm used is a support vector machine, and the unsupervised algorithms are k-means clustering, and affinity propagation, where the latter has not been used before for action recognition. The algorithms are tested on two different datasets. The simple KTH dataset with 6 classes, and the more compli- cated UCF50 with 50 classes.
The SVM classification obtains good results on both KTH and UCF50. Of the two unsu- pervised methods, affinity propagation obtains the best performance. SVM outperforms both of the unsupervised algorithms. The strategy used for building the vocabulary, which is a central part of the bag-of-features framework is tested. The results show, that increasing the number of words will slightly increase the classifier performance.