Spatio-temporal Dynamic Texture Descriptors for Human Motion Recognition - Intelligent Video Event Analysis and Understanding - page 72

to train the non-linear Support Vector Machines (SVM) and the k-Nearest
Neighbors (kNN) classifier.
In the testing phase, an analogous approach is used: for each video, STIPs
are detected and described as feature vectors. Each feature descriptor is suc-
cessively assigned to the closest visual word stored in the vocabulary. The
histogram of spatio-temporal words is then given as input to the classifier
whose output is the class of each video.
As the algorithm has a random component, the clustering phase, any ex-
periment result reported is averaged over 20 runs. The entire methodology
used is shown in Fig. 1

Fig. 1 Methodology

2.2 Feature Detection

Several spatio-temporal feature detection methods have been developed re-
cently and among them we chose Dollar's feature detector [6] because of
its simplicity, fastness and because it generally produces a high number of
responses. The detector is based on a set of separable linear filters which
treats the spatial and temporal dimensions in different ways. A 2D Gaussian
kernel is applied only along the spatial dimensions (parameter to be set),

while a quadrature pair of 1D Gabor filters are applied only temporally (pa-
rameter

σ

τ to be set). This method responds to local regions which exhibit