Multi-object Multi-camea Tracking

Abstract

Multi-camera multi-target tracking aims at following and assigning a consistent label for the moving targets. Tracking has relevance for practical applications; which ranges from medical to security surveillance. However; the behavior of visual data being ambiguous; where targets can be too similar or occluding makes tracking a challenging task.

In this work the challenges of multi-camera multi-target tracking will be considered in a game theoretic approach. For a given objects detection from time synchronized two camera setup, we suggested a method which will use objects in one camera as labels to label objects in the other camera. Assuming objects in each camera is represented as nodes or instances, the basic idea will be followed by the implementation of the relaxation labeling and dominant sets framework to track and assign consistent labels for multiple targets.

In the case of the dominant set framework we build a bipartite graph to show the association between nodes. Then the dominant set is expected to cluster the appearance of the same target in different frames from both cameras with in one cluster. The Relaxation labeling framework used the equivalent symmetrizing idea of the previous framework; by taking objects in both camera as labels to perform labeling the targets in either of the cameras. Both implementation is able to work in on-line and off-line modes.

For the frameworks to decide the similarities and differences between the individual targets; the input datasets are generated based on a probabilistic approach. It is also required that the association graph to base on a robust similarity measure. Using the input data-set from the detection algorithms, we have introduced alternative techniques for tracking and consistent labeling of multi-camera multi-target tracking, which will be an input to the state of the art in the area. The techniques are discussed based on different experiments using toy and standard data-sets, the results are promising and proved the methods are successful for tracking and consistent labeling.

The final contribution of the thesis is an extensive discussion on the performance and comparison with others work; using the standard evaluation metrics for multi-camera multi-object tracking. It concludes by pointing out limitations and future works.