A Non-Parametric Framework for Object Tracking in Videos with Quasi-Stationary Backgrounds
StatisticsView Usage Statistics
The ability to automatically detect and track objects of interest in a video sequence is an essential feature of many high-level vision-based applications. In this dissertation we propose a non-parametric computational framework which detects and tracks foreground image regions (typically corresponding to moving objects such as people or vehicles), while estimating their trajectories for further processing stages such as activity recognition.The contribution of this dissertation extends along two main directions: background modeling for object detection and appearance-based object tracking. In many applications such as vision-based surveillance or traffic monitoring it is typically assumed that the camera is static. Even with this assumption, quasi-stationary backgrounds (that change due to moving tree branches, waving flags, rain or water surfaces) pose significant challenges in detecting and tracking the actual objects of interest while ignoring such changes. Due to the diverse nature of vision-based applications, it has been a main concern for researchers to design a scene-independent system that consistently handles these difficult situations. In the object detection stage, we first propose a novel adaptive statistical method as a base-line system that addresses the issue of scene-independent background modeling. After investigating its performance, we introduce a universal statistical technique which aims to overcome the weaknesses of its predecessor in modeling slow changes in the background. Furthermore, a new analytical technique is proposed that addresses the limitations of statistical techniques which are bound to the probability density estimation accuracy. The performance of each proposed method is studied, while investigating scenarios where each technique leads to better performance.In the object tracking stage, photometric and geometric appearance models are built for the detected objects, then used to predict their locations in subsequent images by employing a novel spatio-spectral tracking technique. The proposed tracker is shown to be particularly robust to local illumination changes, while also being able to detect and resolve the temporary overlapping of tracked objects. We support our claims with extensive experimental results and comparisons between the proposed techniques and other methods for detection and tracking. Finally, we describe a robotic application which successfully employs the proposed vision-based tracking framework in order to infer the intentions of agents in the monitored environment, before their actions are finalized.