Moving Object Detection In Surveillance Video

Moving object detection in surveillance video is a hot research topic in computer vision, which has been widely applied to a variety of applications, such as traffic surveillance and intelligent robotics. Though much progress has been made in recent years, it remains still a challenging problem in complex environments. To this end, this thesis focuses on the global appearance consistency modeling and semantic scene modeling to alleviate the effects of noises and dynamic ground in detecting moving objects. The main works of this thesis are as follows:(1) This thesis proposes a novel method for moving target detection, which pursues the global appearance consistency to improve its robustness. First, we employ the conventional method, ViBe, to obtain the initial detection results, and then estimate the Gaussian Mixture models (GMM) for both foreground and background using these results. Second, we further classify all pixels into foreground or background based on the estimated appearance models. Finally, the superpixel-based refinement is leveraged to obtain the final detection results to alleviate the effects of noises. The experiment result shows that the proposed method obtains superior results against other methods when facing the effects of noises and dynamic background.(2) This thesis proposes a moving target detection method that integrates the spatial distribution and semantic scene priors. First, we utilize some surveillance videos to obtain the spatial distribution information of moving objects, which can reflect the occur probabilities of moving objects. Second, we obtain the semantic information by annotating the scene image. Finally, we integrate these two priors into the GMM based background modeling method to improve the robustness. The experiment results show that this method can eliminate the interference of noises caused by waving trees and meadow, and achieve robust detection results.