Real-time traffic flow topology sensing in partial vehicular ad hoc network: a deep learning solution

In the near future, road traffic is expected to be a mixture of CAVs (connected and autonomous vehicles) and HDVs (human-driven vehicles). CAVs can be monitored by the system. By contrast, the information of HDVs is not readily available. Since full traffic flow topology information is valuable for applications such as data routing, this study proposes a method to infer the information of HDVs using CAV trajectories in partial VANET (Vehicular Ad Hoc Network). Specifically, two deep networks, i.e. VexSen and VelSen, are constructed to infer the existence and location of HDVs. Then, the networks are combined and fused into one sensing procedure. The proposed method was validated by conducting numerous experiments. The results show that HDV existence inferrence accuracy can reach 80%, whereas for the case of multiple CAVs at least the information of 20% HDVs can be inferred given the CAVs penetration rate of 20%.