|
|
1.INTRODUCTIONThe digitalization of transportation facilities and participants is one of the most important development directions of future intelligent transportation systems. The digital twin system can record, reproduce and even interpret the physical system, which helps to promote the digital evolution of the transportation system and the continuous improvement of autonomous driving. It is currently being widely used in intelligent transportation and smart cities1-4. The currently constructed twin traffic systems are more for simulation, and lack the twin practice of rendering real-time models of vehicles and distinguishing vehicle information. When the traffic flow is large, the construction of a real traffic digital twin system puts forward higher requirements for data acquisition, transmission, processing, model calling, rendering, and driving. Currently, there are few high-fidelity real-time traffic twin systems. This paper explores a simple collaborative scheduling method of multi-way side edge computing node resources and the accurate invocation and exit mechanism of the traffic participant model. In order to ensure that a large number of traffic flows and people suddenly appear in some road sections, the amount of data increases, and the fast processing and cloud twin rendering are realized when the roadside nodes have insufficient computing power. Based on the real-time data collected by the large-scale deployment of roadside perception devices (cameras, radars, etc.) on digital roads, combined with AI recognition, a dynamic model library is established. It provides a solution for the rapid twin presentation of mixed traffic and the real-time mapping of traffic participants, traffic events, traffic environment and other elements in the metaverse in the future real world. It can precipitate data and provide scenarios and data foundations for virtual training of autonomous vehicles in the metaverse. 2.A TRAFFIC DIGITAL TWIN SCENE CONSTRUCTION SOLUTION BASED ON EDGE-CLOUD COLLABORATION2.1Technical solution architectureWith the continuous maturity of the digital twin system, comprehensive digital modeling of road pavement, surrounding environment, roadside equipment, etc. can be realized by means of point cloud scanning and aerial photography. And use the sensors distributed on the road to fully perceive and dynamically monitor the traffic conditions. Data is processed by roadside computing nodes and rendered in the cloud, combined with digital maps, to achieve accurate mapping and information expression of physical traffic in the twin world, so as to achieve complete connection between physical space and virtual space5,6. This solution mainly includes three parts: roadside collection terminal, edge node and regional cloud, as shown in Figure 1. 2.1.1Terminal.The roadside collection terminal mainly includes sensors such as cameras and radars, which are responsible for the collection and return of real-time traffic information. 2.1.2Edge Node.The roadside edge node is characterized by miniaturization. It is mainly responsible for the aggregation and processing of a section of road and intersection information, including data access, data analysis, data forwarding, resource monitoring, communication and other modules to achieve collaborative analysis and forwarding of traffic information.
2.1.3Regional Cloud.It mainly includes modules such as roadside node data access, roadside edge node management, various model databases, data processing and model calling, digital twin engine, etc., to realize the construction of real traffic twin scenarios.
2.2Key technologies for the solution2.2.1.Collaborative Data Processing of Multi-way Side Nodes Based on Resource Monitoring.When the traffic flow peaks, the amount of roadside data increases, and the single edge node may have a processing bottleneck, resulting in the inability to upload valid data in real time and affect the construction of twin traffic scenarios. This paper proposes a scheme to efficiently detect the resource usage of edge nodes and allocate and use the resources of surrounding roadside nodes to ensure the efficient processing and uploading of traffic data. This scheme combines active monitoring of node resources with dynamic monitoring that dynamically adjusts the monitoring frequency according to resource usage. Reducing the monitoring frequency that is too low or too high will result in inaccurate monitoring, waste of resources, and the generation of massive monitoring data to run on system resources. The details are as follows:
2.2.2Detection and Identification Technology Based on Multi-source Data Fusion.Aiming at the problem of traffic information acquisition and identification error in bad light at night or in weather such as wind, snow, rain and fog. Through the fusion of radar and video information, the influence of multiple factors can be effectively reduced, and the target detection and recognition of traffic participants can be accurately realized7. Firstly, the radar and video data are preprocessed, and their detection results are tracked respectively, and then the spatiotemporal coordinates of the multi-source data are unified through coordinate transformation and time alignment. Then, through the deep learning model, the information and data from multiple sensors are automatically analyzed and synthesized under certain criteria, and the required decision-making and estimation are finally completed. The main solution used by the deep learning model is the neural network. A large number of simple processing units are fused into a complex model, a large amount of data is imported into the neural network for training, and data features of different dimensions are learned, thereby eliminating the interference of non-target noise, thereby eliminating the noise problem caused by too many sensors. Deep learning models have strong fault tolerance and adaptive capabilities, and can simulate complex nonlinear mapping. For example, for vehicle target recognition, the data information contained in the vehicle can be processed to obtain target feature information. Due to the noise caused by multi-sensor data, the fusion process of these uncertain information is equivalent to the uncertainty inference process. The neural network classifies and deduces the received information on the weight distribution of the network through the acquired information of each sensor, and uses a specific learning algorithm to obtain the learned knowledge and obtain the inference mechanism of uncertainty, thereby realizing multi-sensor data fusion recognition ability. Fusion analysis is mainly divided into three types of solutions: data-level fusion, feature-level fusion and decision-level fusion. The data to be processed in data-level fusion are collected under the same type of sensors, so data fusion cannot handle heterogeneous data. The feature-level fusion is used to reflect the attributes of the monitored physical quantities by extracting the feature vectors contained in the collected data, and is a fusion of the characteristics of the monitored objects. Decision-level fusion performs certain discrimination, classification, and simple logical operations according to the data features obtained by feature-level fusion, and makes higher-level decisions according to application requirements, resulting in application-oriented fusion. We deployed an AI analysis module at the roadside edge node using a feature-level fusion identification scheme as shown in Figure 2. FPN network is used for feature extraction and prediction, and AVOD framework is selected as the basic framework of multi-source data fusion algorithm to detect and identify traffic targets. Traffic participant data is obtained, including participant ID, time stamp, target type (car, bus, bicycle, person, etc.), target feature (license plate number, color, vehicle brand, vehicle type etc.), space location, speed, acceleration, angular speed, etc. This data will be used to drive the traffic actor participant of the twin world. 2.2.3Construction Method of Real Traffic Twin Scene.The system imports road environment data and traffic model data, and relies on the digital twin engine for rendering8. All dynamic elements such as the presentation of the environment, the change of the road state, the loading of the vehicle model, and the change of the driving path are driven by data. Using meteorological information and time information, the rendering engine is driven to simulate weather such as day and night, wind, snow, rain and fog under various lighting conditions to achieve a more realistic virtual traffic environment. For traffic participants, target detection and recognition technology is mainly used to achieve the calibration of the same traffic participant (mainly cars, but also people) through the comparison of ID, license plate or other features, and then the participant model is calibrated and assigned. This technology can realize accurate mapping between real traffic participants and the traffic participant model in the twin system, so as to complete the driving and exit of virtual traffic participants, and maintain the consistency of virtual and real scenes. This paper takes a real-time vehicle in a certain road section as an example to introduce its twin presentation and exit. The following traffic participants are mainly vehicles. The specific steps are as follows, also refer to Figure 1.
The model parameters are a parameter matrix, and each row is related information of a traffic participant ID, which is convenient for the same model to correspond to different traffic participants, such as vehicles of the same type with different license plates. Therefore, deleting a row is equivalent to deleting a traffic participant corresponding to the model. This mode can realize the correspondence between virtual participants and actual traffic participants in the traffic twin, and the entry and exit of its model in the twin traffic scene. Based on accurate target detection and recognition technology, this method can be extended to the realization of virtual presentation and driving of traffic participants including pedestrians, tricycles, etc., and promote the realization of more realistic virtual and real traffic scenes. 3.APPLICATION OUTLOOKDigital twins enable transportation systems to migrate from the physical to the digital world. Combined with the rendering engine, the traffic environment is constructed, and through the fusion and driving of various sensor data, the digital expression of physical entities and processes such as motor vehicles, non-motor vehicles, people, pavement elements and roadside elements on the road is realized. More realistic virtual traffic can be widely used in scenarios such as unmanned virtual testing in various environments, analysis and experimentation of traffic problems, remote operation and maintenance of transportation facilities, and development of new transportation vehicles. In various scenarios, the virtual system can be used to set various conditions and parameters that cannot be set in reality for testing and verification, avoiding possible problems and conducting more innovative practices9,10. With the development of related technologies such as network, interaction, AI, simulation, and real-time rendering, digital twins will affect the real world more, and integrate with the real world, and develop towards the development of virtual and real. The digital twin model can generate certain strategies, dynamically control traffic objects, and realize the interaction between virtual and real traffic. The results of the effect on real traffic will be fed back in the form of data, so as to further optimize the model, realize the autonomous evolution of the model, and drive the transportation development and improvement in the physical world. 4.CONCLUSIONThis paper mainly studies the solution of constructing real-time traffic twin scenes of roads. Based on the collaborative data processing technology of multi-way side edge nodes, it reduces the risk that the data volume of the road section is too large to be quickly processed and uploaded during the traffic peak, and ensures the timeliness of the twin presentation of real-time traffic scenes in the virtual world. A method is proposed to use the AI detection and recognition technology of roadside multi-source data to quickly select the traffic participant model to present, drive and exit in the metaverse. This solution can be widely used to realize real digital twin scenarios of mixed road traffic conditions, and build a digital twin analysis and experimental environment for real traffic operation conditions around the pain points and technological bottlenecks in the transportation industry. In the future, high-quality rendering engines will make virtual scenes more realistic, and AI big data will drive virtual and real traffic to form a closed loop, realize autonomous optimization of the traffic network, and promote the formation of a traffic metaverse where the virtual and the real are merged. REFERENCESWu, Z. H., Wu, X. B. and Wang, L.,
“Prospect of development trend of smart transportation under the background of building China into a country with strong transportation network,”
Transportation Research, 5
(4), 26
–36
(2019). Google Scholar
Tao, F., Liu, W. R. and Liu, J. H.,
“Digital twin and its application exploration,”
Computer Integrated Manufacturing System, 24
(1), 4
–21
(2018). Google Scholar
He, Y., Guo, J. and Zheng, X.,
“From surveillance to digital twin: challenges and recent advances of signal processing for industrial Internet of Things,”
IEEE Signal Processing Mag, 35
(5), 120
–129
(2018). https://doi.org/10.1109/MSP.79 Google Scholar
Tao, F., et al.,
“Digital twin in industry: State-of-the-art,”
IEEE Trans. Industrial Informatics, 15
(4), 2405
–2415
(2019). https://doi.org/10.1109/TII.9424 Google Scholar
Wang, F., Qin, R., Li, J., Yuan, Y. and Wang, X.,
“Parallel societies: A computing perspective of social digital twins and virtual-real interactions,”
IEEE Trans. Comput. Soc. Syst, 7
(1), 2
–7
(2020). https://doi.org/10.1109/TCSS.6570650 Google Scholar
Wang, Z., Liao, X., Zhao, X., Han, K., et al.,
“A digital twin paradigm: Vehicle-to-cloud based advanced driver assistance systems,”
in 2020 IEEE 91st Vehicular Technology Conference,
(2020). https://doi.org/10.1109/VTC2020-Spring48590.2020 Google Scholar
Chen, X., Ma, H. and Wan, J.,
“Multi-view 3D object detection network for autonomous driving,”
in 2017 IEEE Conf. on Computer Vision and Pattern Recognition (CVPR),
(2017). https://doi.org/10.1109/CVPR.2017.691 Google Scholar
Javaid, M. and Haleem, A.,
“Virtual reality applications toward the medical field,”
Clinical Epidemiology and Global Health, 8
(2), 600
–605
(2020). https://doi.org/10.1016/j.cegh.2019.12.010 Google Scholar
Almeaibed, S., AL-Rubaye, S., Tsourdos, A., et al.,
“Digital twin analysis to promote safety and security in autonomous vehicles,”
IEEE Communications Standards Magazine, 5
(1), 40
–46
(2021). https://doi.org/10.1109/MCOMSTD.011.2100004 Google Scholar
Wang, Z., Han, K. and Tiwari, P.,
“Augmented reality-based advanced driver-assistance system for connected vehicles,”
in 2020 IEEE International Conference on Systems, Man, and Cybernetics,
(2020). https://doi.org/10.1109/SMC42975.2020 Google Scholar
|
