Abstract
This study focuses on the problem of object detection in video sequences. In order to free from the modeling of the objects that may be variable and numerous, we choose an a-contrario approach, based on the modeling of the absence of any object of interest in the scene (noise model) and the images are analyzed in order to detect the events salient relative to the used noise model. To free from the a-priori parameters, we define a non-monotonic measure that is a Number of False Alarms (NFA) and that allows quantifying the 'exceptional' feature of an observation relatively to the noise model. With such an approach the solution is determined as the argument minimizing the NFA criterion. Considering an image segmentation problem, the proposed noise model has two, namely pixel and window, levels in order to take into account both ra- diometric and spatial noise feature. The same algorithm is then used to detect either the changes relative to a background, or in the case of an embedded cam- era, the objects having their own motion, or the objects protruding relatively to the road plane, so that applications are related to video surveillance and Advanced Driver Assistance Systems.
Published Version (
Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call