THEORETICAL ADVANCES Adding a rigid motion model to foreground detection: application to moving object detection in rivers Imtiaz Ali • Julien Mille • Laure Tougne Received: 29 October 2011 / Accepted: 12 July 2013 Ó Springer-Verlag London 2013 Abstract Object detection in a dynamic background is a challenging task in many computer vision applications. In some situations, the motion of objects can be predicted thanks to its regularity (e.g., vehicle motion, pedestrian motion). In this article, we propose to model such motion knowledge and to use it as additional information to help in foreground detection. The inclusion of object motion information provides a measure for distinguishing moving objects from a background that has similar sizes and brightness levels. This information is obtained by applying statistical methods on data obtained during the training period. When available, prior knowledge can be incorpo- rated into the foreground detection process to improve robustness and to decrease false detection. We apply this framework to moving object detection in rivers, one of the situations in which classic background subtraction algo- rithms fail. Our experiments show that the incorporation of prior motion data into background subtraction improves object detection. Keywords Mixture of Gaussian  Background subtraction  Object detection  Motion model 1 Introduction Object detection is a widely used task in computer vision applications such as video surveillance, human behavior recognition and video retrieval. In the case of a static camera, background subtraction techniques are used for object detection in many applications. In this kind of approach, a pixel-wise probabilistic representation of the static scene is computed and each input frame is compared to this representation. The mismatch between the two is computed and thresholded so that the moving pixels in the input image are extracted [1–4]. Low computational costs and no requirements of prior knowledge of target objects are two prominent features which make this technique widely popular in the computer vision community. The approach is efficient when the scene to be modeled refers to a static structure with limited perturbations [5]. Moving objects are usually detected according to par- ticular features (i.e., color, motion or shape). Background subtraction methods typically use color/brightness infor- mation alone and do not use object motion characteristics or a priori knowledge of the object shape in foreground or background classification [6, 7]. Pixel-wise difference in background subtraction techniques often leads to mis- classification of background pixels as foreground objects especially in dynamic background conditions. In an aquatic environment, for example, the perturbations in the back- ground brightness are very large. Due to these perturba- tions, scene segmentation is very difficult. In some applications the object motion can be known a priori (e.g., luggage on conveyor belts in airports, vehicle motion, I. Ali (&)  L. Tougne Universit ´ e Lyon 2, Universit ´ e de Lyon, CNRS, LIRIS, UMR5205, 69676 Lyon , France e-mail: engrimtiaz_ali@yahoo.com L. Tougne e-mail: laure.tougne@liris.cnrs.fr I. Ali Optics Labs, PO 1021, Islamabad, Pakistan J. Mille Universit ´ e Lyon 1, Universit ´ e de Lyon, CNRS, LIRIS, UMR5205, 69622 Lyon, France e-mail: julien.mille@liris.cnrs.fr 123 Pattern Anal Applic DOI 10.1007/s10044-013-0346-6