GPU Based Bag of Feature for Fast Activity Detection in Video

Classification of an image on the basis of independent patches selected from image has been effectively used due to its performance. These Image blocks information is collectively represented by Bag of features. Bag of features consist four major parts, block sampling, descriptor generation from collection, characterization of distributions extracted from descriptor and classification of images. In this paper we present GPU based bag of feature generation for effective activity detection in video. Proposed framework focuses on parallel implementation strategy to reduce time taken by traditional bag of visual feature approaches. In bag of visual feature approach clustering takes significant amount of time. We propose GPU based implementation of centroid calculation and dataset generation by distance calculation to reduce time taken by clustering. Feature extraction from video dataset is performed by using MHI with energy and further HOG descriptor. We have used two datasets UT interaction and ATM to validate our results. Our experiments demonstrate that calculations are reduced to almost ten times without affecting accuracy.