PTV Vissim: Paving the road to self-driving technology

Tongji University in Shanghai, China

Estimated Reading Time: 4 minutes

Thanks to the technology behind PTV Vissim, the microscopic simulation software is capable of simulating autonomous vehicles depicting their own motion characteristics and their interaction with other connected or non-connected vehicles. We speak to Prof. Dr. Jian Sun from the Department of Traffic Engineering at the Tongji University in Shanghai, China about leveraging the use of PTV Vissim in his research about Traffic Simulation, Traffic Flow Theory, Intelligent Traffic System.

 

What is the use case to use deep learning with a traffic simulation like PTV Vissim?

Traffic on public roads is complex. The autonomous vehicles must handle all kinds of traffic, road and weather conditions. Otherwise, there will be a significant risk to use autonomous vehicles. Therefore, to ensure safety, training and testing is the principal method to ensure safe autonomous driving on public roads. Among all the testing methods, the virtual test based on the simulation environment (such as PTV Vissim) is considered as the crucial one, for its advantages of lowering costs and shortening developing cycle. Combine the AI algorithms and features of traffic simulator, cases with two aspects are provided.

One is using trajectories from traffic simulator to train the decision-making behaviour of automated vehicles. Deep learning algorithms are widely used on autonomous driving, while the heavy workload on real-world data collection and labelling has constrained the practical application. We propose the idea of training on trajectories from traffic simulator, which is almost zero-cost. It has been implemented in one of the most challenging tasks, the left-turn decision in a two-phase intersection, which could be found in training cases provided by us.

Considering that the deep learning only learns to the offline data, thus the generalisation ability is limited. So, the reinforcement learning is adopted in the other cases, which could not only solve the problem of an insufficient generalisation but also train the driving policy even surpass human driver. The real-time interactive simulation environment is an essential part of the reinforcement learning, which is the advantage of the traffic simulator. Two cases are provided with reinforcement learning; one is the car-following+ lane changing model in road section, the other is the training on the left-turn decision in a two-phase intersection.

Prof. Dr Jian Sun, Department of Traffic Engineering, Tongji University

Who will benefit from your solution?

It’s important for automotive companies especially for that department of autonomous driving, the high-fidelity traffic simulation plays an essential role in the development of driving strategies. It will also benefit AI companies and research institutes; this solution is promising to provide a prototype framework of AI training with the traffic simulation environment so that the trained AI algorithms are more practical. Also, PTV Vissim serves as a powerful traffic simulator which can provide various traffic scenarios and abundant evaluation indexes, so our solution is also available for test organisations of autonomous vehicles.

What scenarios can be modelled with this software set-up (deep learning, PreScan, PTV Vissim)?

Based on our software set-up (the PreScan can provide environment perception information, PTV Vissim can produce background traffic flow, deep learning algorithms are used for decision-making training), various scenarios could be modelled. This is the advantage of PTV Vissim. Since multiple built-in models (e.g. car-following/lane-changing/merging etc.) are available, to reproduce realistic traffic flow, users only need to input the flow volume. While in some existing vehicle simulator, the trajectories need to be added one by one, which wastes a lot of time meanwhile interactions between vehicles are absent. A few ADAS virtual testing tools claimed that they were equipped with the “similar” function as PTV Vissim. However, their simulation on vehicle movement and interaction are not realistic.

Can your solution be used by other companies or be integrated into their software environment?

Yes, our solution can be integrated into another software environment. We use PTV Vissim as the simulator in our training cases and realise the two-way transmission between surrounding vehicles and automated As long as the interface of the software is equipped with corresponding read-write function, the solution is transferable. Users could contact the PTV Group or PTV China to access our cases and related instructions.

Can we replicate local traffic behaviour in the simulation?

Yes, PTV Vissim has been widely used in different countries all around the world. It is applied to simulate filed traffic flow and evaluate traffic situations in many areas, though local traffic may differ from one place to another. In fact, different driving behaviours in the PTV Vissim show as various parameters of the driving behaviour In order to realistically replicate local driver behaviour, the following three steps should establish the simulation model: 1) Data collection before modelling, including data required for model input and model calibration; 2) Establish a high-precision basic model (such as the road network/management and control measures/traffic demand etc.); 3) Model calibration and validation, that is selecting the best-performed combination of driving behavior parameters based on several traffic operational index, to produce the most realistic traffic flow.

 

Want to learn more about PTV Vissim? Click here or download your 30 days free trial now!