Object Detection with Vocabularies of Space-time Descriptors

Abstract

This paper presents a novel framework for objects detection in security and broadcast videos. Our method assumes thatobject classes are unknown in advance and exploit the temporal-space properties of the videos for the creation of avocabulary that describes these classes. Local space-time features have recently became a popular video representationfor action recognition and object detection. Several methods for feature localization and description have been proposedin the literature and promising recognition results were demonstrated for a number of action classes.In this work we propose the use of different kinds of descriptors for the creation of vocabularies for different detectionobject task. For a better description of the videos we carry out a background model, tryring to clean up and follow theareas where there are objects. The points of interest in the videos to characterize the objects are calculated with atemporary variant of the famous Harris corner detector. With the descriptors obtained from the points of interest, avocabulary is realized usingthe kinds of videos we want to train. Then we obtained the frequency histogramsbetween the videos for training and the vocabulary so, with a binary classifier obtain the trained classes and followingthe same procedure without the vocabulary realized the detection and monitoring of the objects.The new method presented is also compared with a state of the art method, obtaining better results in both accuracyand false object rejection.