Target tracking in open world scenes using motion cues and target dynamics

This paper describes the development of a system that will track multiple man made objects (targets), typically vehicles moving in a natural open world scene. The system uses image data at two different resolutions to perform the tracking. Firstly statistical analysis is performed on small areas (tiles) of the original image and inter-frame statistical differences provide motion cues into the original image that define regions of interest. Edge detection is performed within the regions and analysis of the edge information provides an initial indication as to whether the region could be an object or a target. This is repeated on a frame by frame basis with dynamic constraints being applied to both objects and targets. After a number of consecutive frames have been processed, the targets, (objects that have been repeatedly identified and satisfied the dynamic constraints) are identified and displayed on a tracking plot. Objects failing to satisfy the dynamic constraints are removed and no longer processed. Targets that fail to satisfy the dynamic constraints are re-labelled as objects