Learning structural change for identification and tracking of vehicles moving in open world scenes

This paper describes a system, which is able to track multiple man made objects, typically vehicles moving in a natural open world scene. The system interprets changes in structural features in the scene across consecutive image frames and uses the change in structure to identify and track man made objects (target). No a priori knowledge of any structure within the image is assumed. Differences in statistics between a reference image and the current image generate motion cues that are used to identify regions of interest (ROI). Intensity and edgel information are extracted for the ROI in both the current and reference images. Correlation between the extracted data for each ROI yields an initial recognition of the ROI as either a target or object. This process is repeated on a frame by frame basis generating sets of object and target dynamics. The extracted dynamics are used in conjunction with a high-level reasoning process to solve the frame to frame correspondence process. Identified targets are labelled and tracked, objects that fail both the recognition and correspondence process are removed and no longer processed