Samples Based Background Modeling For Object Motion Detection In Video

Video Surveillance systems in cities aim to increase the security and safety feeling among the citizens, and ensure the operability in various civil and military purposes. Researchers tend to reach more developed systems to merge the video analysis of the video feed online.The large amount of data available for the increased number of cameras was accompanied by many research work in object detection topics. Background subtraction forms a basic yet efficient way to detect objects in a video scene.Background subtraction is an old research topic in computer vision. Researchers have investigated the background subtraction algorithms in various occasions, yet there is no perfect universal method that succeeds in managing all the challenges facing such task. A simple search for the contributions in one database leads to a large number of published work on this matter. The importance given to such a step and the various challenges met, encourages researchers to continue the investigation around this hot topic in computer vision. With new theories, a complex modeling, and hybrid representations, one may think about making a global modeling to the problem of background subtraction. However, the application related dependency and the type of video processes makes it an unbalanced problem. These relationships between the application and the modeling choice keep being a key point in selecting a data representation. That is said, an approach for engaging the preprocessing step for a greater goal: Behavior analysis, event understanding, and summarization/categorization.Whereas, a single image could represent a pseudo-static state of observed area, more parametric methods tend to embrace a small agitation of the intensity values around the pseudostatic state, so that further processing allows more considered predicates. Others, rely simply in modeling the background using successive observations, which include only a sampled value from the set. These methods do not imply the real static state restitution rather than having an insight of the slightest changes that may occur. In this thesis, we have considered a study of background subtraction algorithms adopting samples based modeling approach, for what we have called a pseudo-static state of the scene which, named in most of the literature by the “background”. In this thesis, and as a purpose of the presented study, we have considered the accuracy and the robustness to noise of the proposed methods. Aiming to solve these problems, and the research status about object segmentation and background subtraction, the thesis main achievements and innovations included the following:1. Proposed a diversity metric as an adaptive threshold for the classification process foreground/background with short and long frame differences.Various algorithms dealing with background subtraction and foreground/background classification derive their high-speed performance from a fixed threshold value through all the conducted tests. However, researchers aim to replace the decision-making threshold from a static forms into adaptive computation in order to readjust to the real changes that may occur, especially those ignored when using a fixed threshold value. Adopting a fixed value for the classification process could promote the precision to more than 90%, without regard to carry on a distinctive detection of the object. We propose a diversity metric that translates the supposed coherence of the background samples, and as an adaptive threshold value for the classification step. Furthermore, and in order to ascertain the accurate detection of moving objects, a long-term background model represented as a single frame is considered in a three frames’ difference, which consists on the Hadamard product between a short term and a long term frame differences.2. Proposed a samples based modeling approach for moving object detection in video scene.Numerous researchers have proposed formal samples based model of the background model, these approaches considered all the samples to be equally probable in the classification stage. In contrast to such practice, we aimed to propose a weighted process that considers the unbalanced contribution of each sample in the decision-making process, since the farthest sample to the real background value could make the decision-dependent parameters fail, occasionally, an exact interpretation. We showed that the use of proportional weights enhance the accuracy for moving objects detection through an adequately tuned way.3. Proposed a weighted update process in the samples based approaches for the task of Background Subtraction.In the literature, algorithms considering the sample modeling have tendency to update the oldest observation within the background model and keep only the most recent, while others try to do the opposite, so that only the most stationary samples are kept in the background model. Others, recently, have proposed a random selection of the pixels to endure an update.In contrast to these considerations, we proposed an update policy that is non selective nor random. Our method suggested a weighting scheme for the update of the background samples,so that all the background samples are subject to adjustment with regard to their contributions in the belonging decision.4. Proposed a validation process applied to learn the background model coherence for an exact foreground extraction.Pixel-based approaches have tendency to update the background model, regardless the background model’s incoherence that may result, notably, the spatial consistency between neighboring pixels when assuming that the update phase included such characteristic. However, verifying the model coherence could be an efficient way for conserving the uniformity between the background model and the real background scene. We proposed an unsubstantial validation/verification process that learn the steadiness of the background samples.