Intelligent Vehicle Dynamics and Controls (IVDC) – Q&A with Professor Bing Zhu

Events

2

June 2020

WATCH THE FULL REPLAY

The answers below are from Professor Bing Zhu, State Key Laboratory of Automotive Simulation and Control at Jilin University in China, who gave a presentation on Personalised Intelligent Vehicle Control with Driver Behaviour Identification.

How to distinguish among different drivers?

We use many different methods to classify and identify different driver behaviours, such as statistical analysis, machine learning, bionic pattern recognition and so on.

The classification was usually conducted using statistical metrics, such as the means, deviations and maxima of velocity, brake pressure and throttle opening. However, because of the inherent complexity of human driving, it is difficult to measure the similarity between drivers using statistical metrics or subjective indicators based on questionnaires. By analysing the dataset, we found that the distributions of drivers’ driving data significantly differ from each other, and a novel information-theory-based unsupervised clustering method is proposed that can measure the similarity between drivers from the overall driving data distribution. The driving data of every driver is represented as a GMM, then the similarity between human drivers is measured utilising KL divergence.

A skilled driver due to human factors like fatigue may sometimes deviate from the predicted curve. How do you overcome that?

We divide driving behaviour into driving state, driving styles and driving skills. Due to the time, we shared our research on driving skills and driving styles. Driving behaviour is another important factor. And we design the intelligent vehicle controller by these three factors. When the driver is fatigued, the controller will help the driver to maintain the vehicle stable.

At the same time, the drivers in the driver behaviour data acquisition were in normal condition without fatigue or drunkenness. And a safety supervisor was assigned to the test vehicle to ensure safety.

There are too many dimensions to describe the driving behaviour of each driver. Why do you select three types of driving styles?

The proper classification of driving styles is the precondition of personalised vehicle control research. And driver-related factors can be complex and varying, different drivers have different driving behaviours even in the same conditions. If there are too few types of driving styles, the results won’t be accurate.

On the contrary, too many categories of driving styles will make it difficult for personalising vehicle control. It is a classic classification to divide driving styles into three types. In current researches, most scholars classify driving styles into three categories of cautious, normal, and aggressive drivers. It is proved that the classification can describe driving styles in a simple and comprehensive way. At the same time, we are using NN like GRU to learn from data of each driver and aim to develop a personalisation system that can accommodate driver-specific behaviour for vehicle dynamics control.

Do you state that safe driving is correlated with "cautious" driving or in case "aggressive driving" could be safe sometimes?

We don`t think safe driving is directly correlated with “cautious” driving. The concept of safe driving needs to be measured from many aspects. The “cautious” driving and “aggressive” driving discussed in our study are descriptions of different driving behaviour patterns. They can`t reflect on safety alone.

The “cautious” driving describes a modest driving behaviour. For example, a “cautious” driver may start to decelerate at long distance from the front vehicle. The “aggressive” driving doesn`t mean the driver drives rudely or thoughtlessly. It just reflects a behaviour pattern. In contrast to “cautious” driver, a “aggressive” driver may start to decelerate at short distance from the front vehicle.

What is the cost-function defined here? If the driver changes his/her driving style during real-world driving, how does your control strategy to counter this situation and ensure the global optimality?

The cost-function considers four main parts, which are driving behaviour indicator, tracking indicator, comfort indicator, and economic indicator. The driving behaviour indicator makes the longitudinal acceleration decision inclined to imitate the driver's personalised behaviour.

For the adaptability of the personalised control strategy, we have developed online driver identification method and control system dynamic adjustment method. After a specific time-window, the identification method will give a classification result and the control system can adjust its strategy according to the result to ensure that the control model is consistent with the driver’s style.

Can the driver classifications be used to create "virtual drivers / agents" to simulate different behaviour in simulation systems for autonomous driving?

The classification results can be used to generate driver models with different kinds of driving behaviours. Some of the parameters in the driver model are set on the classification. These driver models can react to the environment according to their own behaviour settings. Therefore, the driver models derived from the classification results can be used in simulation environment to simulate driving actions with different behaviour.

Do you develop driver monitoring as fatigue or hypo-vigilance?

Developing driver monitoring system is of great significance. We are now working on driver monitoring system using camera, eye tracker, EMG (Electromyography). etc.

How many driver's feature dimension in the driver model training, and which is the most impact factor influencing the driver's classification?

As mentioned earlier, the classification was usually conducted using statistical metrics, such as the means, deviations and maxima of velocity, brake pressure and throttle opening. However, because of the inherent complexity of human driving, it is difficult to measure the similarity between drivers using statistical metrics or subjective indicators based on questionnaires. By analysing the dataset, we found that the distributions of drivers’ driving data significantly differ from each other, and a novel information-theory-based unsupervised clustering method is proposed that can measure the similarity between drivers from the overall driving data distribution. The driving data of every driver is represented as a GMM, then the similarity between human drivers is measured utilising KL divergence.

Related Links for IVDC Online Conference

WATCH THE FULL REPLAY

IVDC Moderator Summary by Dr Mike Ma, Professor at Jilin University in China and VP Technical, FISITA

Questions answered by Dr Xinjie Zhang following his presentation on Cooperative Adaptive Cruise Control (CACC) Including Platform Kinematics

Questions answered by Dr Ye Zhuang following his presentation on Vehicle Driving Dynamics and Control for Active Safety under Extreme Manoeuvre

Vehicle Dynamics and Controls will continue to evolve as we move ever closer to fully automated driving. This area is just one being considered as part of Intelligent Safety. If you wish to learn more about the subject and continue the conversation, the Intelligent Safety Conference will be held in conjunction with the China SAE Congress & Exhibition (SAECCE), in October 2020, in Shanghai. Please visit the FISITA website www.fisita.com/events/isc for more information.