Efficiency of synthetic vision in simple scenarios

Abstract: Modern crowd simulation models can display a range of different properties, such as realism, efficiency, or emergent self-organizing patterns. These properties often vary by the specific situation that the model is used in. A model can be realistic when used in one scenario and break down completely in another scenario. Knowing the limits of a model is essential to know when it can and cannot be used. Furthermore, describing these limits, and explaining their underlying causes, can allow improving the model, or creating a new model, to surpass the previous model in terms of flexibility of use. Lastly, testing potential changes to a model is important for finding the right way to improve that model in the future. In this study, a synthetic vision-based model by Ondřej et al. is tested. The model attempts to recreate human perception and decision-making for a more realistic navigation algorithm. In the paper where it was proposed, the model showed great efficiency in complex scenarios with a high agent density and high number of agents. This was due to highly efficient self-organizing emergent patterns that allowed the individual agents to reach their goals quickly compared to existing models, such as the RVO model. Ondřej et al. did not test their model in simpler scenarios. We aim to help further inform about which use cases the synthetic vision model is appropriate for, and to lay the groundwork for future efficiency improvements of the synthetic vision model in low-density scenarios. This paper reimplements and tests the synthetic vision-based model by Ondřej et al. in simpler scenarios by comparing its efficiency to the RVO model. We also investigate potential improvement of the synthetic vision model by varying a model parameter. The results show that the model completely fails in one simple test case, and that it is clearly outperformed by RVO in another. One problem revealed by the results is that the model has difficulty with adapting its path far enough ahead of a future collision, and that this problem can be at least somewhat mitigated by varying a model parameter. Two other problems were also revealed. The first is that the agents adjust their paths very weakly towards their goal. The second is that the agents sometimes unnecessarily slow their speed while close to another agent. We describe the underlying cause of each of these three problems based on the mathematical definition of the synthetic vision crowd simulation model. Our contribution should aid others in choosing to use this synthetic vision model in appropriate scenarios, and should be a valuable foundation for improving this model in the future.