Intelligent Surveillance Of Moving Objects In Complicated Traffic Scenes

With the fast social development, public safety has caught more and more attention in all the world. A large number of video surveillance systems are widely deployed in many areas. These traditional surveillance systems, however, are lack of intelligence in the sense that their major function is to record videos without understanding the recorded contents. Such systems suffer from low processing capacity and are not fully utilized. To resolve these issues, intelligent videos surveillance emerges as a new solution, and is the hot topic in academics, industries and administration agencies.This thesis focuses on the traffic video surveillance systems and attempts to solve some problems. Traffic scenes are the most concerned places to monitor. Therefore, our research has a strong application background and the obtained results can be implemented in reality. The traffic scenes are often complicated and challenging. They usually suffer from many kinds of disturbances. The objects in such scenes may be occluded by shadows and obscured by other objects. What’s more, the characteristics of object may change dramatically when the angle and distance between the object and camera change. In these traffic scenes, it is challenging to detect, track and classify objects correctly.Based on the available theories in computer vision and the latest research results on traffic video surveillance systems, we intensively investigate object detection, tracking and classification in the traffic scenes, and propose some approaches to tackle existing problems. In the end, our proposed approaches are implemented to design and realize an intelligent traffic video surveillance system, which can be used to give warnings on object abnormality and retrieve a specified object very fast. The contributions of this thesis can be summarized as follows.1) This thesis proposes a background modeling approach based on the feedback of the tracking results of moving objects. The traditional approaches always adopt a uniform updating strategy to update the background model of the whole scene. However, the characteristics of pixels in the scene are different. The background models of some pixels have to update as soon as possible, while the ones of other pixels should update slowly, and some even need not to update. In this thesis, we propose to utilize the feedback of tracking results for background model updating. The scene is divided into four different kinds of regions, and then different adaptive updating strategies are taken for different types of regions. The proposed approach can well deal with the trade-off between model robustness to background changes and model sensitivity to foreground abnormalities. Moreover, the misclassification of the type of region nearly has no harm to the performance. And the proposed approach has a low complexity, which can meet the needs of real-time application.2) This thesis proposes a foreground segmentation approach based on the feedback of the tracking results of moving objects. The general approaches always adopt a uniform segmentation strategy for the whole scene. However, the characteristics of pixels in the scene are different. In some pixel regions, it has to adopt stringent strategy for foreground segmentation so as to suppress noise, reduce false alarms. While in other pixel regions, a loose strategy should be adopted to reduce erroneous holes and splitting of objects. In this paper, we propose to feed back the tracking results for foreground segmentation. Based on the feedback information, it predicts the object regions in the following frame. In the object regions, adaptive segmentation thresholds are taken for foreground segmentation to reduce erroneous holes and splitting. In the non-object regions, a large threshold is taken for foreground segmentation to reduce false alarms.3) This thesis proposes an occlusion-adaptive multi-object tracking approach. In the traffic scenes, the occlusion among objects and the change of objects’appearance are inevitable. If we can’t well deal with these problems, it will greatly degrade the object tracking performance. According to the spatial-temporal continuity, we divide object motion into three states, including the independence state, the occluded state and the splitting state. And then three different tracking strategies are implemented for corresponding states. The proposed approach can correctly detect the objects’motion state and track the objects, both in the cases of partial occlusion and total occlusion.4) This thesis proposes an adaptive approach for object classification. The general approaches adopt a uniform rule for object classification and can not work well when the characteristics of the objects change with their movement in the scene. Our new approach takes the divide and conquer technique. First, it trains different classification rules for different regions in the scene, and then gives the final decision of the object category by fusing the decision of the object in different regions. Experiments show that the proposed approach can well distinguish pedestrians, bicycles and cars. 5) The above proposed approaches are integrated to design and realize an intelligent traffic video surveillance system. It gives warnings on object abnormality and retrieves a specified object very efficiently.