Video Scene Object Detection Using An A Contrario Approach

This paper aims at showing the interest of the a contrario framework for object detection in video scenes. In the two approaches presented here, objects are detected at a window level, considering all the pixels included in a given window to decide the presence or the absence of objects in the window. Now, according to the a contrario principle, windows with objects are detected as too exceptional realizations of the model representing the windows without objects. The interest of this latter model, called `naive' model, is that it is generally much simpler than the one representing the variety of objects. Two window-based algorithms are proposed, one using the fact that an appearing object can be characterized by significantly high values in the image representing the difference with a background or reference, and the other one using the fact that objects form clusters of object-labeled pixels. The performance of our approach (two algorithms) has been tested on video scenes respectively acquired outdoors and indoors. Both algorithms have also been compared to alternative detection methods, and they proved their performance Finally, the obtained results on artificial noised images show the high robustness relatively to noise of the proposed two-step detection method.