Real-time Ship Track Association: a Benchmark and a Network-Based Method

Multiple systems are tracking ships independently, including cooperative as well as non-cooperative ones. Most track association works focused on association cost functions and optimal track matching matrix was determined by solving a multidimensional assignment problem, assuming synchronous detections of fixed number of targets received periodically from an ideal communication channel, which is not satisfied when tracking targets in a large scale by heterogeneous methods. We put forward an evolving target network-based association approach to create, update, merge, separate and destroy targets over time, which centers on targets rather than tracks. In the network, nodes are fused ships corresponding to tracks from one or more sources and weights of directed edges pointing to each node are association beliefs constituting its basic probability assignment (BPA) function in Dempster-Shafer theory (DST). Once a new message is received, this network evolves by combining the corresponding node's BPA with a new one calculated according to the message. This method is evaluated by an extensible benchmark consisting of an open source track simulator ICTShips and a free visualization software ICTShipVisual. The benchmark consists of four scenes with 32 to 102,231 targets observed by 3 sources at present, having GERs (Gap Error Ratio) from 0.787 to 1,412. Error ratios of our track association method under these scenes range from 0% to 0.539%.