Researchers at the Advanced Digital Sciences Center (ADSC) have proposed two new methods of object tracking, including what they believe to be the first which considers correlation amongst particles. Both methods are based on a particle filter framework and outperform existing methods in terms of robustness and efficiency. Tracking objects in video is a fundamental research problem in the computer vision world. The goal of object tracking in video is to be able to follow the movement of any interesting target, such as a player in a sports video, a face in a busy crowd or a car in a parking garage, over time. Tracking objects is difficult due to occlusion, motion blur or similar appearances of other objects, amongst other reasons.
In ADSC’s first method, the Cross-Domain Tracker (CDT), the researchers use cross-domain contextual information for sports video analysis from both the image domain (video frames, see image below left) and field domain (blueprint of football field, see image below right). The second method, the Multi-Task Tracker (MTT) is completed by mining interdependencies amongst particles to improve tracking performance and reduce computational complexity. ADSC’s CDT method has 30 percent more accuracy than conventional trackers. The second method has a 10 percent accuracy improvement and is twice as fast as the popular L1 tracker, which is closely related to ADSC’s tracker, on benchmark datasets.
The image above shows a frame from an American football video clip (left) and the tracking results (right) of multiple players after CDT tracking. ADSC’s methods can improve the accuracy of object tracking compared to conventional trackers by 30 percent.
ADSC researchers have applied their tracking research primarily to sports videos through the ADSC Semantic Analysis of Video project, which aims to develop a framework for the automated analysis of complex activities present in video. Due to the accomplishments in tracking, ADSC’s researchers are able to track athletes in sports videos, which allows for analysis of specific players, plays and ultimately entire games. In addition to sports analysis, object tracking plays an important role in automatic surveillance, robotics and human computer interaction, among other areas. The object tracking research can be applied in many areas because it allows a viewer to understand the motion of a target. Based on that information, they can recognize events and analyze any video.