Consistent Object Tracking Via Multiple Underwater Cameras

In this paper, we set up a distributed underwater surveillance system that adopts multiple static cameras to track object trajectory. The system consists of an Information Fusion Unit (IFU) as well as a set of Underwater Image Processing Units (UIPU), which can combine and coordinate the tracking tasks consistently among multiple cameras.The tracking process of the UIPU is facilitated in a discriminative appearance model by distinguishing the target and the background with the segmented superpixels in the surrounding region, and then formulated by calculating a target-background confidence map. The location of the target can then be obtained by the maximum a posterior estimation. The UIPU for each camera records the surveillance background, and transfers both the dynamic features and the target labels to the IFU.The IFU detects the overlapping field of views among the cameras and matches the background by the Speed Up Robust Features (SURF) which can describe the interest points that are scale and rotation invariant. SURF relies on on integral image and Hessian matrix to reduce the computation time.The viewpoint correspondence among the overlapping fields can then be established by the homographic transformation matrix, which can be accurately calculated by the Random Sample Consensus (RANSAC) algorithm with key points extracted from SURF. In order to track the target consistently and effectively, the IFU utilizes the centroid coordinate homographic mapping and SURF together to match the same target across the overlapping fields, while only the SURF features are used among the non-overlapping fields. The corresponding labels can be marked for the same tracking target across multiple cameras at last.The simulation experiments show that the scheme proposed in this paper can continuously achieve the consistent target tracking results across multiple cameras in a wide range of the scene under the sea.