| Automated surveillance systems are of critical importance for the field of security. The task of reliably detecting and tracking moving objects in surveillance video, which forms a basis for higher level intelligence applications, has many open questions. Our work focuses on developing a framework to detect moving objects and generate reliable tracks from real-world surveillance video. After setting up a basic system that can serve as a platform for further automatic tracking research, we tackle the question of variation in distances between the camera and the objects in different parts of the scene (object depth) in surveillance videos. A feedback-based solution to automatically learn the distance variation in static-camera video scenes is implemented, based on object motion in different parts of the scene. The solution, based on a concept that we call 'Vicinity Factor', is robust to noise and object segmentation problems. The Vicinity Factor can also be applied to estimate spatial thresholds and object size in different parts of the scene. Further, a new Bayesian algorithm to assign tracks to objects in the video is developed. The Bayesian method allows for probabilistic track assignment and can be the basis for future higher-level inference. |
