Object classification for real-time video-surveillance applications

Object classification is a fundamental step in automatic video-surveillance that allows improved tracking and a more accurate description of events. However, as real-world applications need a real-time, flexible, easy and quick to configure solution, the design of a practical object classification algorithm becomes a challenge. This paper analyses advantages and disadvantages of different frameworks presented in the literature, with particular focus on the ones that are more suitable for real-world operation. A learning-based solution using a reduced training set is proposed, demonstrating that it overcomes many of the limitations associated with other algorithms.