Semantic interpretation of dynamic image sequences attempts to automatically interpret the motions and behaviors of the tracked objects by the analysis of the image sequences captured by cameras from wide-area,real-world scenes in natural conditions.Traditional computer vision research mainly focuses on recovering the geometry of the scene(structure from motion),the camera motion (ego-motion),and the motion of the pixels themselves(such as optical flow).In recent years,semantic interpretation of image and video has become an active research topic in computer vision. In this thesis,we study the semantic interpretation of vehicle motion in traffic scenes which involves many basic problems in computer vision,e.g.motion detection,object localization,spatial-temporal reasoning,behavior analysis and semantic interpretation,etc.The main contributions of this thesis include the following: ①We have proposed a simple but convenient method for camera calibration in traffic scenes,an improved motion detection algorithm with lower sensitivity to lighting,and an efficient and robust vehicle localization algorithm. ②We have described a modified extended Kalman filter with 8 novel kinematics model for visual vehicle tracking.By imposing an additional orthogonality condition,the filter is less sensitivity to the temporal variations of the system model.Experiments show that the filter has a good performance when the tracked car is in complex motion. ③A framework for semantic interpretation of vehicles'motion has been proposed in the context of visual traffic surveillance.We introduce a conceptual space to bridge the gap between quantitative low-level processing and qualitative high-level processing. ④From human's mental experiences,there are two aspects of abstraction: "generality"and"complexity".We deal with them in two different computational stages named"conceptual processing"and"symbolic processing" to simplify the modelling and inference of them. ⑤We have presented a new interval-based model of action and a temporal analvzer to model and recognize the objects'behaviors in traffic scenes A single object’sbehaviors and its interactions with other objects can be handled in the same framework.Finally,some of the recognized actions can be selected and translated into natural language descriptions by some simple grammar rules. ⑥We have developed a demo platform for further research which can work at a speed of 17 frames per second on a computer with PIV 1.7G CPU and Windows operating system.The system can give some simple semantic interpretations of vehicle’s behaviors.
修改评论