Abstract

Recently, automated video surveillance and related video processing algorithms have received considerable attention from the research community. Challenges in video surveillance rise from noise, illumination changes, camera motion, splits and occlusions, complex human behavior, and how to manage extracted surveillance information for delivery, archiving, and retrieval: Many video surveillance systems focus on video object extraction, while few focus on both the system architecture and video object extraction. We focus on both and integrate them to produce an end-to-end system and study the challenges associated with building this system. We propose a scalable, distributed, and real-time video-surveillance system with a novel architecture, indexing, and retrieval. The system consists of three modules: video workstations for processing, control workstations for monitoring, and a server for management and archiving. The proposed system models object features as temporal Gaussians and produces: an 18 frames/second frame-rate for SIF video and static cameras, reduced network and storage usage, and precise retrieval results. It is more scalable and delivers more balanced distributed performance than recent architectures. The first stage of video processing is noise estimation. We propose a method for localizing homogeneity and estimating the additive white Gaussian noise variance, which uses spatially scattered initial seeds and utilizes particle filtering techniques to guide their spatial movement towards homogeneous locations from which the estimation is performed. The noise estimation method reduces the number of measurements required by block-based methods while achieving more accuracy. Next, we segment video objects using a background subtraction technique. We generate the background model online for static cameras using a mixture of Gaussians background maintenance approach. For moving cameras, we use a global motion estimation method offline to bring neighboring frames into the coordinate system of the current frame and we merge them to produce the background model. We track detected objects using a feature-based object tracking method with improved detection and correction of occlusion and split. We detect occlusion and split through the identification of sudden variations in the spatia-temporal features of objects. To detect splits, we analyze the temporal behavior of split objects to discriminate between errors in segmentation and real separation of objects. Both objective and subjective experimental results show the ability of the proposed algorithm to detect and correct both splits and occlusions of objects. For the last stage of video processing, we propose a novel method for the detection of vandalism events which is based on a proposed definition for vandal behaviors recorded on surveillance video sequences. We monitor changes inside a restricted site containing vandalism-prone objects and declare vandalism when an object is detected as leaving the site while there is temporally consistent and significant static changes representing damage, given that the site is normally unchanged after use. The proposed method is tested on sequences showing real and simulated vandal behaviors and it achieves a detection rate of 96%. It detects different forms of vandalism such as graffiti and theft. The proposed end-ta-end video surveillance system aims at realizing the potential of video object extraction in automated surveillance and retrieval by focusing on both video object extraction and the management, delivery, and utilization of the extracted information