Moving Object Detection Based On Adaptive Background Model

Moving object detection (also named change detection) is an important topic in computer vision. It aims at accurately and completely extracting moving objects which are of uses’interest from videos. It always attracts much attention, mainly attributing to two reasons. On one hand, as a mid-level vision part, moving object detection provides basis for high-level vision tasks such as object tracking, object classification and behavior understanding. As a result, it is widely applied in intelligent video surveillance. On the other hand, moving object detection is very challenging. Complex background, dynamic noises, trivial motion that is not of interest for users, and bad weather bring large challenges. To detect moving objects robustly and quickly attracts many researchers. With decades of developments, many detection methods are proposed. However, some theoretical and technical problems still need our further efforts, such as how to accurately model background and how to quickly adapt dynamic changes of scenario. In this thesis, an adaptive background model is proposed. This method can adjust and optimize the background model by analyzing the complexity of scenario. Based on this, it realizes moving object detection from videos. Besides, we incorporate super-pixel based segmentation to optimize the regions of moving objects. Our main work as follows:(1)Based on studying the properties of dynamic noises in videos, we proposed a measure to estimate the dynamic properties of scenario. It extracts dynamic noise regions by exploiting spatial-temporal relations among frames. Correspondingly, a quantitative measure is given to help discriminate moving objects in foreground and moving noises in background. This provides a basis for the following adaptive background model.(2)A moving object detection method based on adaptive background model is proposed. As a pixel-level method, it estimates the complexity of scenario and integrates the dynamic property measure in (1). It has two advantages. One is that it can dynamically adjust our method’s important parameters without user intervention, contrast to traditional methods which use manually adjusted parameters or constants. The other is that the use of pixel-level modeling facilitates high-speed implementations. Our own version, which used no parallel optimization, reaches real-time processing speed. Experimental results on public dataset CDnet2014 show that our method is robust in different types of videos. This method is very close to the best detection methods in terms of overall F-Measure, and runs at a speed of 23fps.(3)A super-pixel based moving region optimization method is proposed. Exploiting the prior that nearby pixels share similar features, we optimize the results from (2). This makes the regions corresponding to moving objects more complete. Experimental results show that this step of optimization can saliently improve detection performance on part of the videos in CDnet2014. In vision tasks where real-time requirement is not so rigid, this approach can provide more accurate object regions.