At present, our country relies on cross-industry coordination mechanisms, infrastructure construction and information and communication technology advantages to promote the development of assisted driving technology through “vehicle-road-cloud integration”. By 2025, more than 22,000 kilometers of test roads have been opened nationwide, more than 5,200 test licenses have been issued, and the total mileage of road tests has reached 88 million kilometers. At the policy level, in 2018, the Ministry of Industry and Information Technology promoted the “Internet of Vehicles (Intelligent networkingIn 2021, the Ministry of Housing and Urban-Rural Development and the Ministry of Industry and Information Technology jointly launched a pilot project of “dual intelligence” cities, and in 2024, five ministries and commissions launched the application pilot of “vehicle-road-cloud integration”, marking that technology verification has entered the commercialization stage.
On July 22, 2025, Guo Yuan, an associate researcher at Wuhan University of Technology, said at the 8th Intelligent Assisted Driving Conference in 2025 that the dynamic cognitive map is based on static high-precision maps, integrating the dynamic information of multiple observation nodes at the vehicle end and the roadside end, and realizing the consistent expression and scene understanding of dynamic and static traffic information, which can significantly improve the safety, energy saving and road adaptability of the assisted driving system, and effectively solve the bottleneck of the development of high-level assisted driving.
Guo Yuan | Associate researcher of Wuhan University of Technology
The following is a summary of the speech:
Research background and core questions
our country’s vehicle-road-cloud integration technology has achieved leapfrog development from proof-of-concept to commercial landing. As of 2025, a total of 22,000+ kilometers of open test roads will cover 16 provinces across the country, 5,200+ test licenses will be issued covering multiple categories of passenger cars/commercial vehicles, and the total road test mileage of 88 million kilometers is equivalent to 2,200 orbits around the earth. In terms of policy drive, the Ministry of Industry and Information Technology’s “Action Plan for the Development of the Internet of Vehicles Industry” established a technical standard system in 2018; In 2021, the “dual intelligence” city pilot will promote the intelligent transformation of infrastructure; In 2024, the pilot application of “vehicle-road-cloud integration” of five ministries and commissions will include vehicle-road collaboration in the scope of national new infrastructure for the first time, marking that the technology has officially entered the stage of commercial deployment. Through dynamic cognitive map technology, the real-time coupling of physical road space and vehicle-side dynamic information is realized, providing faster, more accurate and more comprehensive environmental cognition capabilities for the assisted driving system, and is committed to achieving the goals of safe, energy-saving and efficient operation.
Source: Speaker material
As the core carrier to realize the abstraction of physical space to information space, high-precision maps accurately depict the geometry, attributes and relationship information of road geographical elements on the basis of traditional navigation electronic maps, and their abstraction process is reflected in the transformation of physical roads into structured data models, such as the four-layer integrated model developed by Wuhan University. However, the existing mainstream formats including Shapefile, Occupancy Grid, OpenDrive, etc. still have significant limitations: on the one hand, they focus too much on static road information, and lack effective depiction of dynamic elements such as vehicle trajectories and traffic events. On the other hand, the complex model structure leads to high acquisition costs and difficulty in supporting real-time update requirements, which seriously restricts the real-time response ability of assisted driving systems in complex scenarios.
Source: Speaker material
Dynamic cognitive map model construction
The dynamic cognitive map uses the static high-precision map as the spatial benchmark, and constructs a collaborative cognitive model with a three-layer architecture by integrating the real-time perception data of vehicle-side sensors and roadside units: the underlying static layer inherits the lane-level topology network of the high-precision map to provide a unified spatio-temporal reference for dynamic information; The dynamic fusion layer processes the motion state of multi-source traffic elements in real time and performs them based on the attention mechanismforecastto generate the trajectory probability distribution for the next 5 seconds; The cognitive graph layer analyzes the semantics of traffic scenes through visual language models, and constructs a hierarchical dynamic map of the subordinate regions of roadside units combined with graph neural networks to realize the interpretable reasoning of traffic situations. The architecture supports six core functions: traffic event perception to achieve accident identification, spatio-temporal situation prediction, environmental weather perception, vehicle/vehicle-person collision warning response, signal light phase synchronization, blind spot warning, and finally form a closed loop of assisted driving decision-making of “static bottoming, dynamic stacking, and cognitive driven”.
Source: Speaker material
The dynamic cognitive map model realizes precise support for six core functions through the multi-source spatio-temporal data fusion mechanism: traffic event perception is based on the spatio-temporal correlation between roadside radar and on-board sensors, which can quickly identify traffic accidents, road construction and other events; Spatio-temporal situation prediction uses the attention mechanism to integrate historical trajectories and real-time positions to predict traffic flow changes in the next 5 seconds, with a trajectory prediction error of ≤ 0.8 meters; Environmental weather perception combines millimeter-wave radar echo and visual image analysis to identify rain, fog, ice and snow and other weather; Vehicle/vehicle collision warning calculates the collision time through the relative kinematics model to achieve collision warning. The signal light phase reminder realizes millisecond synchronization between the vehicle end and the roadside signal. Blind spot warning and vehicle speed induction use V2X communication to cover visual blind spots. This functional system breaks through the limitations of traditional static maps, builds a dynamic and static collaborative closed loop of “perception-prediction-decision-control”, and provides full-scenario support capabilities for continuous and reliable lane-level positioning and second-level response for the assisted driving system.
Source: Speaker material
Breakthroughs in key technologies of dynamic cognitive maps
At the level of perception technology innovation, the research team proposes the MENet algorithm, which fuses the geometric structure features of high-precision maps (such as lane curvature and curb topology) with the depth information of laser point clouds by constructing a laser point cloud driving scene perception framework under the unified spatio-temporal benchmark of high-precision maps, so as to solve the problem of scale inconsistency between vehicle perception and local maps. In the public dataset test, the target detection mAP of the algorithm reached 56.9%, which was 4.2% higher than the commonly used deployment baseline CenterPoint, of which the truck recognition accuracy was 52.3% and the bus was 66.2%, significantly optimizing the robustness of multi-object detection at intersections.
Source: Speaker material
The simultaneously developed MIM model (Map-incorporated Multi-view) adopts hierarchical semantic analysis technology to target the semantic scale differences between multi-view images and high-precision maps: firstly, the structured features such as lane markings and traffic signs in the map are extracted through mask augmentation, and then the cross-focus mechanism is combined to realize the cross-modal alignment of image pixel features and map vectorized features. In the end, the target detection mAP reached 40.9% in the nuScenes test, especially reducing the vehicle positioning error at the intersection to less than 0.3 meters.
Source: Speaker material
In the field of map matching, lane-level positioning is realized based on the lateral displacement estimation model, with a lateral error ≤ 0.3 meters, which meets the centimeter-level requirements of the assisted driving system for lane keeping. The large-scale point cloud registration technology RoCalib extracts the feature clouds of light poles, signs and other poles in high-precision maps, and matches the features with roadside radar scan data to solve the problem of roadside equipment pose drift, and the time synchronization accuracy reaches milliseconds, providing spatio-temporal reference guarantee for dynamic cognitive maps.
Source: Speaker material
Future challenges and development directions
The current technology faces three core challenges: the difficulty of perception fusion is manifested in the presence of millisecond-level synchronization errors in the spatio-temporal reference alignment of multi-source heterogeneous data, and the complexity of multimodal information fusion leads to a deviation rate of up to 15% for dynamic target recognition; Dynamic tracking is difficult to concentrate in low penetration scenarios, and the difficulty of small target recognition detection reduces the over-the-horizon continuous tracking accuracy to 65%, especially for non-motor vehicle blind spot trajectory prediction failure risk is significant. The difficulty of expression optimization stems from the surge in the amount of dynamic information data, and the imbalance of model training data leads to the limitation of accurate expression of map information.
The key research directions in the future include: multi-source information coupling technology in mixed transportation scenarios,Lightweightdynamic map expression model, dynamic cognitive architecture migration based on ordinary navigation electronic map, and optimization of over-the-horizon tracking algorithm in low-penetration environment. Research and verification show that these breakthroughs can reduce the cost of the assisted driving system by 60% and improve the adaptability of complex urban road conditions to 99.5% reliability.
(The above content comes from the keynote speech on “Research on Key Technologies for the Construction of Vehicle-Road-Cloud Integrated Dynamic Cognitive Maps” delivered by Associate Researcher Guo Yuan of Wuhan University of Technology at the 8th Intelligent Assisted Driving Conference in 2025 on July 22, 2025.) ）
