摘要：Public benchmark datasets have been widely used to evaluate multi-target tracking algorithms. Ideally, the benchmark datasets should include the video scenes of all scenarios that need to be tested. However, a limited amount of the currently available benchmark datasets does not comprehensively cover all necessary test scenarios. This limits the evaluation of multitarget tracking algorithms with various test scenarios. This paper introduced a computer simulation model that generates benchmark datasets for evaluating multi-target tracking algorithms with the complexity of multitarget tracking scenarios directly controlled by simulation inputs such as target birth and death rates, target movement, the rates of target merges and splits, target appearances, and image noise types and levels. The simulation model generated a simulated video and also provides the ground-truth target tracking for the simulated video, so the evaluation of multitarget tracking algorithms can be easily performed without any manual video annotation process. We demonstrated the use of the proposed simulation model for evaluating tracking-by-detection algorithms and filtering-based tracking algorithms.
