英文摘要 | Visual traffic surveillance is to perform recognition, localization and tracking of road vehicles based on video sequences captured by cameras without human intervention or with little human intervention, and further analysis vehicles' behaviours according to the tracking results. It has important potential applications in the area of traffic management and surveillance for particular scenarios, such as highway entrance and exit, toll station, dangerous section, parking lot, residential quarter, etc. In the past decade, visual traffic surveillance has been a popular issue in computer vision with increasing concern. However, it has not reached the maturity for large-scale applications yet, while the primary limitation lies in three aspects: precision, processing speed, robustness. In this thesis, we present a number of novel algorithms for road vehicle localization and tracking from monocular image sequences, which is a fundamental problem in visual traffic surveillance. Based on these algorithms, we have designed and implemented a prototype vehicle tracking system, which is also introduced in the end of the thesis. The contributions of this thesis include: 1) A novel 3D wireframe model based vehicle localization algorithm is presented. It considers localization process as a series of virtual motions, which can be decomposed into two kinds of independent motions: translation and rotation. The motion parameters for both of them can be determined in closed form. Experiment results show that it can efficiently and robustly derive vehicles' 3D pose from one intensity image with desirable precision. 2) An improved Extended Kalman Filter (EKF) based vehicle tracking algorithm is presented. Considering some physical characters of moving vehicles, we introduce a new dynamic model for their motion. The novelty of our dynamic model is that the internal structures of vehicles are taken into account, while most existing algorithms consider moving vehicles as particles. Our algorithm also utilize an improved EKF, which incorporates the orthogonality condition into traditional EKF to counteract the instability of model parameters, thus satisfying the EKF's assumption of measurement noise's being White Noise. Experiment results show that it can obtain better performance than traditional EKF under complicated driving maneuver. 3) A novel partial match based occlusion reasoning approach is presented. It can automatically adjust the model to be matched with the image when occlusion occurs. We divide occlusions into three kinds, which include vehicles occluded by the background, vehicles occluded when entering or leaving the view of camera, vehicles occluded by other vehicles, and present processing strategies for all of them. Experiment results show that it can still obtain goo |
修改评论