Article Preview
Top1. Introduction
Video surveillance system is an important part of Intelligent Transportation System (ITS). Detecting and tracking moving objects from video sequences is one of the important tasks in video surveillance system. Recent days, many approaches are proposed in this field. In Betke, Haritaoglu, and Davis (2000) and Lin and Xu (2006) a combination of color, edge, and motion information is used to recognize and track the road boundaries, lane markings and other objects on the road. Cars are recognized by matching templates that are cropped from the input data online and by detecting highway scene features and evaluating how they relate to each other. But the method is difficult in handling significant scale changing of the object, and threshold is subjectively determined with less robustness. In Sun and Zhu (2008), after getting the contours of moving objects, region growing segmentation method is applied to locate the foreground region, and the rectangle centre is seen as the mass centre of the moving objects. Moving objects are tracked using minimal Euclidean distance and the motion trajectories are also painted. However, it performances bad when the centroids of objects are close to each other, and also it is subject to shadow. So, when in complex environment, the accuracy will be greatly reduced. Kalman filtering is used in Rowe, Reid, Gonzàlez, and Villanueva (2006) combined with block-based color histogram matching for multi-target tracking. However, in complex environment, initial parameters can not be appropriately obtained, which often leads to accumulation of tracking error. In (Comaniciu & Ramesh, 2000; Comaniciu, Ramesh, & Meer, 2003) a mean-shift method is presented for objects tracking. Mean-shift method manifests high efficiency for target tracking with low complexity. But as a hill climbing algorithm, it may fall into local most low and lose the motion target when occlusion occurs. In Okuma, Taleghani, Freitas, Little and Lowe (2004) and Cai, Freitas, and Little (2006) a boosted particle filtering combined with Mean-shift method is proposed for multi-object tracking; it can effectively track the human bodies and other non-rigid objects, and is robust for mutual occlusion. However, it requires a large number of samples for prior training, and the computational complexity is high, which limits its application.
A 3D-model-based vehicle tracking has been proposed in Ottlik and Nagel (2008) and Dahlkamp, Nagel, Ottlik, and Reuter (2007), which is based on Edge-Element and Optical-Flow association. It tracks vehicles by fitting the models to the vehicle boundaries. However, the ambiguity in determining vehicle boundary will cause the tracking accuracy reduced when mutual collusion occurs or in complex background. Also, if the wrong vehicle type has been chosen, or if the object candidate does not cover most of the vehicle image due to scale changing drastically, the result is bad. In (Ross, Lim, Lin, & Yang, 2008; Ross, Lim, & Yang, 2004), an efficient and effective online method is presented which incrementally learns and adapts a low dimensional eigen- space representation to reflect appearance changes of the target, and then to facilitate the tracking task. The method is robust to large pose and lighting changes. However, it needs many images to learn eigenspaces under different views for representation and tracking, which limits the application in un-surveillance or on-line tracking.