I have got a video containing 200 frames and i am extracting low level features at pixel level. Then i am dividing the entire frame in to small spatial patches (say of size 7x7) and creating a feature vector for each spatial patch (by summing up the feature values of the pixel inside the particular patch). If i want to use the HoG features in the above mentioned procedure, what is the best possible way to do that? i.e using PCA? or shall i use the values of the bin for that particular 7x7 patch directly (say if i have 4 bins then it will be 4 feature variables)?
PS : i am using GMM to model the features. So i want the features of each patch can be correlated with one-another using euclidean distance measure.