Classification of Compressed Videos

Abstract

Classification of Compressed Videos is one of the challenging aspect in today’s date. Analysing and classifying the contents of videos is an important factor in retrieval of data. Around 72,000 videos are uploaded on youtube per minute so it becomes very difficult to classify all the videos manually or using certain low level features only. In order to properly classify the data content of a video, we need to have a proper knowledge of the various features of a video. Previously, many classification tasks has been performed based on the text,audio,video or many a times a fusion of them like HMM uses both text and audio. Algorithms like Bag of visual words which is used for action recognition is used on dataset UCF50 for classifying all the different 50 classes in their respective domain and viola jones method is used for human detection and recognition. Features like motion vectors, gradient descent and optical flow are extracted from the video frame and provided as an input to SVM for classification. Maximum accuracy in results is achieved when different modalities are combined together so that maximum of the features gets extracted.