Online training of object detectors from unlabeled surveillance video

One of the decisive steps in automated surveillance and monitoring is object detection. A standard approach to constructing object detectors consists of annotating large data sets and using them to train a detector. Nevertheless, due to unavoidable constraints of a typical training data set, supervised approaches are inappropriate for building generic systems applicable to a wide diversity of camera setups and scenes. To make a step towards a more generic solution, we propose in this paper a method capable of learning and detecting, in an online and unsupervised setup, the dominant object class in a general scene. The effectiveness of our method is experimentally demonstrated on four representative video sequences.