Service area dynamic parking guidance method based on real-time trajectory perception
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WUHAN METROPOLITAN RING ROAD NORTHERN SECTION INVESTMENT MANAGEMENT CO LTD
- Filing Date
- 2026-02-04
- Publication Date
- 2026-05-29
AI Technical Summary
The existing highway service area parking system suffers from functional failure of static guidance mechanism, lack of behavioral perception capability, and failure of parking space allocation strategy to integrate vehicle movement intention characteristics, resulting in mismatch between resource allocation and demand, causing traffic congestion within the site.
A dynamic parking guidance method based on real-time trajectory perception is adopted. By fusing GNSS/IMU positioning and roadside visual tracking, combined with dual-stream LSTM spatiotemporal trajectory prediction and MADDPG multi-agent collaboration, a vehicle-type sensitive real-time path is generated, and hard-constrained path execution is implemented to optimize parking space allocation.
It improved parking lot operation efficiency, enhanced positioning accuracy and resource utilization, reduced the missed detection rate in low light/rain and fog scenarios, ensured accurate allocation of charging spaces for electric vehicles, and reduced traffic congestion.
Smart Images
Figure CN122108180A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent transportation technology, and in particular to a method for dynamic parking guidance in service areas based on real-time trajectory perception. Background Technology
[0002] Currently, most parking lots in China share three common characteristics: low traffic efficiency, difficulty in finding parking spaces, and delayed information access. When drivers enter a parking lot, they often cannot obtain real-time status information in a timely manner. Quickly locating available parking spaces in service area parking lots is quite difficult and highly unpredictable, directly leading to low overall parking lot operating efficiency, a need to improve service quality, and frequent traffic congestion.
[0003] The current highway service area parking system is facing two core dilemmas. First, the static guidance mechanism is functionally ineffective, as traditional fixed signs and one-way information transmission modes are difficult to adapt to the needs of dynamic scenarios. Second, there is a systemic lack of behavioral perception capabilities. The existing system has significant blind spots in the real-time capture of driver operation characteristics, and the current parking space allocation strategy fails to fully integrate the movement intention characteristics of vehicles in the ramp area. Specifically, it fails to establish a quantitative evaluation model for the probability of decelerating into the area, and it does not dynamically capture key operational characteristics such as lane-changing tendencies, ultimately resulting in a structural mismatch between resource allocation and actual needs.
[0004] Therefore, a service area dynamic parking guidance method based on real-time trajectory perception is proposed to solve the above problems. Summary of the Invention
[0005] The purpose of this invention is to address the aforementioned shortcomings by providing a dynamic parking guidance method for service areas based on real-time trajectory perception.
[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a service area dynamic parking guidance method based on real-time trajectory perception, including a dynamic guidance system, wherein the dynamic guidance system specifically includes a trajectory perception layer, a prediction and decision layer, and a path planning layer. The trajectory perception layer integrates vehicle-mounted GNSS / IMU positioning with roadside visual tracking; The prediction decision layer is based on spatiotemporal trajectory prediction using dual-stream LSTM combined with MADDPG multi-agent collaboration. The path planning layer is used for vehicle-type sensitive real-time path generation; The specific process is as follows: S101, multi-source real-time data acquisition and fusion positioning, through GNSS / IMU hardware combination, adopts third-order extended Kalman filter to dynamically compensate sensor zero bias, eliminates multipath effect caused by building occlusion, deploys improved visual perception module, and combines attitude correlation cost function to solve target loss in rain and fog occlusion scene, outputting a real-time vehicle position matrix with centimeter-level accuracy. S102, Spatiotemporal trajectory prediction and behavior modeling: Input absolute coordinates and the distance to the nearest obstacle, construct geometric constraints, capture speed changes and heading angular rate, generate temporal features, dynamically aggregate dual-stream features through self-attention weight α, embed service area acceleration and deceleration laws, and force the predicted trajectory to conform to the motion law; S103. Hierarchical collaborative decision optimization generates a dynamic map, triangulates the set of vacant parking spaces, filters out obstacle edges, and constructs a centimeter-level occupancy matrix. S104. Hard constraint path execution and feedback: Perform vehicle type-sensitive allocation based on real-time congestion index. Update the topological edge weights, feed back the actual vehicle trajectory to the perception layer, and iteratively optimize the prediction model.
[0007] Furthermore, a multi-source fusion sensing module is used to achieve real-time data acquisition and fusion positioning from multiple sources. The vehicle-mounted terminal adopts a third-order extended Kalman filter architecture to establish a state vector. Through the pre-integral model Separate motion acceleration from sensor zero bias; The visual perception module employs a roadside vision-deployed cascaded detection framework, specifically including: Improve YOLOv7 by introducing a lightweight MobileNetV3 backbone and high-resolution feature layers to enhance small target detection; To optimize DeepSort, an incremental feature library update strategy is designed, which involves using the association cost function. Solve the problem of ID obscuring.
[0008] Furthermore, a dual-stream spatiotemporal prediction module is used for spatiotemporal trajectory prediction and behavior modeling, and a multi-agent collaborative decision-making module is used for hierarchical collaborative decision optimization. This multi-agent collaborative decision-making module is based on a three-objective optimization function. The multi-agent collaborative decision-making module is equipped with a hierarchical decision-making mechanism; The trajectory perception layer constructs a topological grid bimodal map and dynamically updates the obstacle distribution; The MADDPG algorithm in the prediction decision layer defines the continuous action space. ; Improvements in the path planning layer Algorithm, in heuristic function Introducing vehicle model factors .
[0009] Furthermore, a hard constraint mechanism is used to implement hard constraint path execution and feedback. This hard constraint mechanism is as follows: if... For electric vehicles and Allocate charging spaces; if Charging stations are prohibited for vehicles that are fuel-powered.
[0010] Furthermore, the specific process of the GNSS / IMU hardware combination includes: S201. Establish a time reference between the GNSS receiver and the IMU module to achieve second-level synchronization and initialize the dual-buffered queue system. S202. Perform real-time acquisition of dual-source data. The GNSS module continuously receives satellite data and simultaneously acquires motion parameters. S203. Perform cross-sensor trajectory matching, calculate the distance matrix of the two trajectories respectively, and reverse the optimal path to calculate the time offset.
[0011] Furthermore, in terms of spatial modeling, the GNSS / IMU hardware combination converts IMU accelerometer data to the navigation system using multi-dimensional state vectors, as shown in the formula: ,in, Satellite elevation angle, parameters , Calibration was performed using cluster analysis of ground reflection characteristics in the service area; The GNSS / IMU hardware combination uses extended Kalman filtering to fuse GNSS pseudorange observations and IMU inertial data, and introduces dynamic compensation for sensor bias. The GNSS / IMU hardware combination is designed with an incremental feature library dynamic update strategy. This strategy continuously optimizes the feature representation space, as expressed in: ; in, Cosine distance of apparent features; For vehicle posture similarity; The weighting coefficients are determined by actual measurement to obtain the optimal values.
[0012] Furthermore, the dual-stream spatiotemporal prediction module includes: S301: Temporal and spatial flow inputs, feature encoding and modeling; S302: Perform feature concatenation, calculate spatial weights, and perform weighted fusion; S303: Apply physical law constraints and output prediction results; Establish spatial flow, input absolute position and obstacle distance, and model geometric constraints. The spatial flow expression is: ;in, The change in vehicle speed For the heading angular rate, For time step; A timeline is established to capture velocity changes and heading angular rate, and temporal evolution is modeled. Adaptive fusion is performed, and features are dynamically aggregated through self-attention weights. The time flow expression is: ;in, Indicates the absolute position of an object. Distance to the nearest obstacle; The dual-stream spatiotemporal prediction module employs a cascaded visual localization framework, using YOLOv7 target detection and DeepSort for multi-target tracking. The adaptive optimization expression for YOLOv7 is as follows: ;in, For classification confidence; The focus coefficients for difficult samples are obtained by introducing MobileNetV3 to replace the original convolutional layer, which significantly reduces the computational cost. A small target detection layer is used to optimize the weights of difficult samples.
[0013] Furthermore, the dual-stream spatiotemporal prediction module employs an adaptive fusion mechanism, which dynamically adjusts the weights of information including vehicle position and distance to surrounding obstacles with information including speed changes and heading angular rate, thereby enabling the automatic allocation of spatial and temporal features based on the real-time scene using a self-attention mechanism. The expression for the adaptive fusion mechanism is: ; in, For the future The predicted state vector at each time step. and Here are the weight matrix and bias vector of the fully connected layer; Physical constraints are constructed in the dual-stream spatiotemporal prediction module; The physical constraint expression is: ,in, The distance between the target vehicle and the vehicle in front; This is the critical safety distance; This is the Gamma distribution function.
[0014] Furthermore, the information input section of the dual-stream spatiotemporal prediction module is a parking space status perception layer, which includes an vacant parking space topology map. and dynamic occupancy matrix The spatial coordinates of the parking space are defined as follows: ,in, East-west coordinates, North-south coordinates Indicates the orientation of the parking space.
[0015] Furthermore, regarding the aforementioned The algorithm is improved by introducing a dynamic adjustment factor into the heuristic function, expressed as follows: ,in, From the starting point to The actual cost, for Estimated cost to the destination, dynamic adjustment factor It should meet the following requirements: .
[0016] The beneficial effects of this invention are: centimeter-level positioning accuracy is achieved in complex environments through multi-source fusion; the false negative rate of the vision module is reduced in low-light / rainy / fog scenes; and the hierarchical decision framework reduces the complexity of path planning from... Down to This effectively improves resource utilization, and the hard constraint mechanism of charging piles effectively improves the matching accuracy of electric vehicles. Attached Figure Description
[0017] Figure 1 This is a schematic diagram of the dynamic guidance system of the present invention; Figure 2 This is a flowchart of the GNSS / IMU hardware assembly of the present invention; Figure 3 This is a flowchart of the dual-stream spatiotemporal prediction module of the present invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] Please see Figure 1 This invention discloses a dynamic parking guidance method for service areas based on real-time trajectory perception, including a dynamic guidance system. The dynamic guidance system specifically includes a trajectory perception layer, a prediction and decision-making layer, and a path planning layer. The trajectory perception layer integrates vehicle-mounted GNSS / IMU positioning with roadside visual tracking; The prediction decision layer is based on spatiotemporal trajectory prediction using dual-stream LSTM combined with MADDPG multi-agent collaboration; The path planning layer is used for vehicle-type sensitive real-time path generation; The specific process is as follows: S101, multi-source real-time data acquisition and fusion positioning, through GNSS / IMU hardware combination, adopts third-order extended Kalman filter to dynamically compensate sensor zero bias, eliminates multipath effect caused by building occlusion, deploys improved visual perception module, and combines attitude correlation cost function to solve target loss in rain and fog occlusion scene, outputting a real-time vehicle position matrix with centimeter-level accuracy. S102, Spatiotemporal trajectory prediction and behavior modeling: Input absolute coordinates and the distance to the nearest obstacle, construct geometric constraints, capture speed changes and heading angular rate, generate temporal features, dynamically aggregate dual-stream features through self-attention weight α, embed service area acceleration and deceleration laws, and force the predicted trajectory to conform to the motion law; S103. Hierarchical collaborative decision optimization generates a dynamic map, triangulates the set of vacant parking spaces, filters out obstacle edges, and constructs a centimeter-level occupancy matrix. S104. Hard constraint path execution and feedback: Perform vehicle type-sensitive allocation based on real-time congestion index. Update the topological edge weights, feed back the actual vehicle trajectory to the perception layer, and iteratively optimize the prediction model.
[0020] In this embodiment, multi-source fusion sensing module is used to achieve real-time data acquisition and fusion positioning from multiple sources. The vehicle-mounted terminal adopts a third-order extended Kalman filter architecture. Establish state vector , Through the pre-integral model Separate motion acceleration from sensor zero bias; The visual perception module employs a roadside vision-based cascaded detection framework, specifically including: Improve YOLOv7 by introducing a lightweight MobileNetV3 backbone and high-resolution feature layers to enhance small target detection; Optimize DeepSort by designing an incremental feature library update strategy. Through the correlation cost function Solve the problem of ID obscuring.
[0021] In this embodiment, a dual-stream spatiotemporal prediction module is used for spatiotemporal trajectory prediction and behavior modeling, and a multi-agent collaborative decision-making module is used for hierarchical collaborative decision optimization. The multi-agent collaborative decision-making module is based on a three-objective optimization function: The multi-agent collaborative decision-making module includes a hierarchical decision-making mechanism; A topological grid bimodal map is constructed in the trajectory perception layer to dynamically update the obstacle distribution; The MADDPG algorithm defines a continuous action space in the predictive decision layer. ; Improvements in the path planning layer Algorithm, in heuristic function Introducing vehicle model factors .
[0022] In this embodiment, hard constraint path execution and feedback are implemented through a hard constraint mechanism. The hard constraint mechanism is as follows: if For electric vehicles and Allocate charging spaces; if Charging stations are prohibited for vehicles that are fuel-powered.
[0023] The trajectory perception layer employs a fusion solution of vehicle-mounted GNSS / IMU positioning and roadside visual tracking. Combined with a third-order extended Kalman filter algorithm to dynamically compensate for sensor bias, it effectively eliminates multipath effects caused by building obstruction in service areas, significantly improving positioning stability and accuracy. Ultimately, it outputs a real-time vehicle position matrix with centimeter-level accuracy, providing reliable data support for subsequent trajectory prediction and path planning. Simultaneously, the visual perception module improves upon YOLOv7 by introducing a lightweight MobileNetV3 backbone and a high-resolution feature layer, enhancing small target detection capabilities. Coupled with an optimized DeepSort and incremental feature library update strategy, it leverages an association cost function to address target loss and ID switching issues in rain and fog obstruction scenarios. This ensures the system maintains high-efficiency perception performance even in adverse weather and complex environments, adapting to diverse operating scenarios in service areas.
[0024] Please see Figure 2 In this embodiment, the specific process of combining GNSS / IMU hardware includes: S201. Establish a time reference between the GNSS receiver and the IMU module to achieve second-level synchronization and initialize the dual-buffered queue system. S202. Perform real-time acquisition of dual-source data. The GNSS module continuously receives satellite data and simultaneously acquires motion parameters. S203. Perform cross-sensor trajectory matching, calculate the distance matrix of the two trajectories respectively, and reverse the optimal path to calculate the time offset.
[0025] In this embodiment, the GNSS / IMU hardware combination converts IMU accelerometer data to the navigation system through multi-dimensional state vectors in spatial modeling, as shown in the formula: ,in, Satellite elevation angle, parameters , Calibration was performed using cluster analysis of ground reflection characteristics in the service area; Although a single IMU can continuously output pose data, its zero bias error increases quadratically over time, so a dynamic error compensation mechanism needs to be established. The GNSS / IMU hardware combination uses extended Kalman filtering to fuse GNSS pseudorange observations and IMU inertial data, and introduces dynamic compensation for sensor bias. The expression is: ; in, , For vehicle plane coordinates; , For velocity components; For heading angle; , To achieve zero bias in the accelerometer; Zero bias for the gyroscope; Separating motion acceleration from sensor bias using a pre-integral model: ; in, This is the raw acceleration of the IMU; Let be the acceleration due to gravity, and take . ; Based on traditional ReID feature extraction, an incremental feature library dynamic update strategy is designed in the GNSS / IMU hardware combination; Specifically, when the vehicle trajectory is not occluded within 3 consecutive valid frames, the system uses a dynamic weighted average strategy to update its historical appearance features; when the occlusion is removed, the updated feature library is used for appearance similarity matching. This strategy, through continuous optimization of the feature representation space, effectively solves most identity switching problems caused by temporary license plate obscuring. The expression is: ; in, Cosine distance of apparent features; For vehicle posture similarity; The weighting coefficients are determined by actual measurement to obtain the optimal values.
[0026] Please see Figure 3 In this embodiment, the specific workflow of the dual-stream spatiotemporal prediction module includes: S301: Temporal and spatial flow inputs, feature encoding and modeling; S302: Perform feature concatenation, calculate spatial weights, and perform weighted fusion; S303: Apply physical law constraints and output prediction results; A timeline is established to capture velocity changes and heading angular rate, and temporal evolution is modeled. Adaptive fusion is performed, and features are dynamically aggregated through self-attention weights. The time flow expression is: ;in, Indicates the absolute position of an object. Distance to the nearest obstacle; Establish spatial flow, input absolute position and obstacle distance, and model geometric constraints. The expression for space flow is: ;in, The change in vehicle speed For the heading angular rate, For time step; To address issues such as abrupt changes in illumination and visual distortion in parking lot environments, the dual-stream spatiotemporal prediction module employs a cascaded visual localization framework, using YOLOv7 target detection and DeepSort for multi-target tracking. The adaptive optimization expression for YOLOv7 is: ;in, For classification confidence; The focus coefficients for difficult samples are obtained by introducing MobileNetV3 to replace the original convolutional layer, which significantly reduces the computational cost. A small target detection layer is used to optimize the weights of difficult samples.
[0027] In this embodiment, the dual-stream spatiotemporal prediction module adopts an adaptive fusion mechanism, which dynamically adjusts the weights of information including vehicle position and distance to surrounding obstacles with information including speed changes and heading angular rate, so as to realize the automatic allocation of spatial and temporal features according to the real-time scene using a self-attention mechanism. The expression for the adaptive fusion mechanism is: ; in, For the future The predicted state vector at each time step. and Here are the weight matrix and bias vector of the fully connected layer; The behavior of vehicles in the service area can be divided into acceleration and deceleration phases. During the acceleration phase, the vehicle acceleration follows a Gamma distribution, while during the deceleration phase, the vehicle deceleration is hyperbolic and tangential to the distance between vehicles. Therefore, physical constraints are constructed in the dual-stream spatiotemporal prediction module. The physical constraint expression is: ,in, The distance between the target vehicle and the vehicle in front; This is the critical safety distance; The Gamma distribution function; Simultaneously, path feasibility constraints are constructed, and a regularization term is added to the loss function: ,in, To predict acceleration; To predict the heading angle; Lane direction angle; , The regularization coefficient is used. Based on this, priority constraints are implemented to prevent charging spaces from being occupied by gasoline vehicles: ;in, The number of charging stations available in real time; When an electric vehicle enters the service area parking lot, if there are still available charging parking spaces, the system enters the normal scheduling process and selects the optimal target from the available parking spaces; when an electric vehicle enters but there are no available charging piles or a fuel vehicle attempts to be assigned a charging pile, the route allocation terminates.
[0028] In this embodiment, the information input part of the dual-stream spatiotemporal prediction module is the parking space status perception layer, which is responsible for transforming the physical parking lot into a computer-understandable digital model. Its core task is to extract two key features from the raw sensor data. The parking space status perception layer includes an vacant parking space topology map. and dynamic occupancy matrix The spatial coordinates of the parking space are defined as follows: ,in, East-west coordinates, North-south coordinates Indicates the orientation of the parking space; To integrate scattered parking space status data into structured information, enabling rapid filtering of invalid data and improving system efficiency, the following data structure is defined: ,in, This indicates the parking space status value: 0 means the space is available for parking, 1 means the space is occupied, and -1 means the space is unavailable due to a malfunction. This refers to the total number of parking spaces in the service area parking lot. Construct a topology map of vacant parking spaces The topology map describes the reachable path relationships between vacant parking spaces, providing a road network foundation for route planning; The expression is: ,in, This is a set of vertices, corresponding to each available parking space; ;in, For the vertex spatial position, start from the parking space coordinates where the state is 0. The reason for choosing this value is that only vacant parking spaces will become navigation target points; Next, perform triangulation and define a triangular mesh set. for: Among them, the adoption The subdivision satisfies the "maximum and minimum interior angle criterion" to ensure that the generated triangles are as close to equilateral as possible; A single triangular unit is a planar triangle formed by three vacant parking spaces in space; To prevent vehicles from having undesirable driving trajectories, feasible edges for obstacle filtering are constructed. ; ,in, The initial edge set generated by the triangulation; Let the set of obstacles be such that if the edges When an edge intersects with any obstacle, that edge is removed. The initial triangular edges may represent routes that are not actually traversable. It is necessary to ensure that all edges in the graph are physical paths that vehicles can travel on. In order to improve traffic efficiency and prevent congestion, dynamic weight allocation is implemented. The expression is: ;in, This is the Euclidean distance, i.e., the basic path cost; A dynamic congestion factor adjusted based on real-time traffic flow; Since continuous coordinates are not conducive to rapid collision detection, a dynamic occupancy matrix is constructed to reflect the real-time spatial occupancy of vehicles. A grid-based approach is used to describe the overall occupancy status, which can serve as the core basis for collision detection. A grid space mapping matrix is defined. : ; Parking space coordinates are mapped using the principle of mapping. Convert to raster index; ,in, The dimensions of the raster matrix; Raster resolution; The visual detection should follow the rule that a vehicle occupies the grid. The geomagnetic sensor reports whether a parking space is occupied or faulty. season Finally, the perception results are output. To represent the tensor of the topological graph's occupancy matrix, we define the state space: ;Yaw adjustment of guidance equipment, defining the motion space: ,in, For time step The action vector, for Yaw angle adjustment of the guidance equipment; To reduce the communication overhead of full information sharing, a two-level attention structure is designed: ; in, , These are the projection matrices of the query vector and the key vector, respectively; For equipment A set of neighboring devices within a fixed radius, enabling each device to dynamically monitor the status of its most relevant neighboring devices, reducing communication load from... Down to .
[0029] In this embodiment, for The algorithm is improved by introducing a dynamic adjustment factor into the heuristic function, expressed as follows: ,in, From the starting point to The actual cost, for Estimated cost to the destination, dynamic adjustment factor It should meet the following requirements: ; satisfy:
[0030] in, This refers to the width of the accessible passageway in the parking lot. For based on Updated real-time congestion index.
[0031] In summary, this invention achieves centimeter-level positioning accuracy in complex environments through multi-source fusion, reduces the false detection rate of the vision module in low-light / rainy / foggy scenes, and lowers the path planning complexity from... Down to This effectively improves resource utilization, and the hard constraint mechanism of charging piles effectively improves the matching accuracy of electric vehicles.
[0032] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A service area dynamic parking guidance method based on real-time trajectory perception, characterized in that, It includes a dynamic guidance system, which specifically comprises a trajectory perception layer, a prediction and decision-making layer, and a path planning layer. The trajectory perception layer integrates vehicle-mounted GNSS / IMU positioning with roadside visual tracking; The prediction decision layer is based on spatiotemporal trajectory prediction using dual-stream LSTM combined with MADDPG multi-agent collaboration. The path planning layer is used for vehicle-type sensitive real-time path generation; The specific process is as follows: S101, multi-source real-time data acquisition and fusion positioning, through GNSS / IMU hardware combination, adopts third-order extended Kalman filter to dynamically compensate sensor zero bias, eliminates multipath effect caused by building occlusion, deploys improved visual perception module, and combines attitude correlation cost function to solve target loss in rain and fog occlusion scene, outputting a real-time vehicle position matrix with centimeter-level accuracy. S102, Spatiotemporal trajectory prediction and behavior modeling: Input absolute coordinates and the distance to the nearest obstacle, construct geometric constraints, capture speed changes and heading angular rate, generate temporal features, dynamically aggregate dual-stream features through self-attention weight α, embed service area acceleration and deceleration laws, and force the predicted trajectory to conform to the motion law; S103. Hierarchical collaborative decision optimization generates a dynamic map, triangulates the set of vacant parking spaces, filters out obstacle edges, and constructs a centimeter-level occupancy matrix. S104. Hard constraint path execution and feedback: Perform vehicle type-sensitive allocation based on real-time congestion index. Update the topological edge weights, feed back the actual vehicle trajectory to the perception layer, and iteratively optimize the prediction model.
2. The service area dynamic parking guidance method based on real-time trajectory perception according to claim 1, characterized in that: The multi-source fusion sensing module enables real-time data acquisition and fusion positioning from multiple sources. The vehicle-mounted terminal employs a third-order extended Kalman filter architecture to establish a state vector. Through the pre-integral model Separate motion acceleration from sensor zero bias; The visual perception module employs a roadside vision-deployed cascaded detection framework, specifically including: Improve YOLOv7 by introducing a lightweight MobileNetV3 backbone and high-resolution feature layers to enhance small target detection; To optimize DeepSort, an incremental feature library update strategy is designed, which involves using the association cost function. Solve the problem of ID obscuring.
3. The service area dynamic parking guidance method based on real-time trajectory perception according to claim 2, characterized in that: Spatiotemporal trajectory prediction and behavior modeling are performed through a dual-stream spatiotemporal prediction module, and hierarchical collaborative decision optimization is performed through a multi-agent collaborative decision module, which is based on a three-objective optimization function. The multi-agent collaborative decision-making module is equipped with a hierarchical decision-making mechanism; The trajectory perception layer constructs a topological grid bimodal map and dynamically updates the obstacle distribution; The MADDPG algorithm in the prediction decision layer defines the continuous action space. ; Improvements in the path planning layer Algorithm, in heuristic function Introducing vehicle model factors .
4. The service area dynamic parking guidance method based on real-time trajectory perception according to claim 3, characterized in that: Hard constraint path execution and feedback are implemented through a hard constraint mechanism, wherein the hard constraint mechanism is: if For electric vehicles and Allocate charging spaces; if Charging stations are prohibited for vehicles that are fuel-powered.
5. The service area dynamic parking guidance method based on real-time trajectory perception according to claim 1, characterized in that: The specific process for combining the GNSS / IMU hardware includes: S201. Establish a time reference between the GNSS receiver and the IMU module to achieve second-level synchronization and initialize the dual-buffered queue system. S202. Perform real-time acquisition of dual-source data. The GNSS module continuously receives satellite data and simultaneously acquires motion parameters. S203. Perform cross-sensor trajectory matching, calculate the distance matrix of the two trajectories respectively, and reverse the optimal path to calculate the time offset.
6. The service area dynamic parking guidance method based on real-time trajectory perception according to claim 5, characterized in that: In terms of spatial modeling, the GNSS / IMU hardware combination converts IMU accelerometer data to the navigation system using multi-dimensional state vectors, as shown in the formula: ,in, Satellite elevation angle, parameters , Calibration was performed using cluster analysis of ground reflection characteristics in the service area; The GNSS / IMU hardware combination uses extended Kalman filtering to fuse GNSS pseudorange observations and IMU inertial data, and introduces dynamic compensation for sensor bias. The GNSS / IMU hardware combination is designed with an incremental feature library dynamic update strategy. This strategy continuously optimizes the feature representation space, as expressed in: ; in, Cosine distance of apparent features; For vehicle posture similarity; The weighting coefficients are determined by actual measurement to obtain the optimal values.
7. The service area dynamic parking guidance method based on real-time trajectory perception according to claim 3, characterized in that: The dual-stream spatiotemporal prediction module includes: S301: Temporal and spatial flow inputs, feature encoding and modeling; S302: Perform feature concatenation, calculate spatial weights, and perform weighted fusion; S303: Apply physical law constraints and output prediction results; Establish spatial flow, input absolute position and obstacle distance, and model geometric constraints. The spatial flow expression is: ;in, The change in vehicle speed For the heading angular rate, For time step; A timeline is established to capture velocity changes and heading angular rate, and temporal evolution is modeled. Adaptive fusion is performed, and features are dynamically aggregated through self-attention weights. The time flow expression is: ;in, Indicates the absolute position of an object. Distance to the nearest obstacle; The dual-stream spatiotemporal prediction module employs a cascaded visual localization framework, using YOLOv7 target detection and DeepSort for multi-target tracking. The adaptive optimization expression for YOLOv7 is as follows: ;in, For classification confidence; The focus coefficients for difficult samples are obtained by introducing MobileNetV3 to replace the original convolutional layer, which significantly reduces the computational cost. A small target detection layer is used to optimize the weights of difficult samples.
8. The service area dynamic parking guidance method based on real-time trajectory perception according to claim 7, characterized in that: The dual-stream spatiotemporal prediction module employs an adaptive fusion mechanism, which dynamically adjusts the weights of information including vehicle position and distance to surrounding obstacles with information including speed changes and heading angular rate, thereby enabling the automatic allocation of spatial and temporal features based on the real-time scene using a self-attention mechanism. The expression for the adaptive fusion mechanism is: ; in, For the future The predicted state vector at each time step. and Here are the weight matrix and bias vector of the fully connected layer; Physical constraints are constructed in the dual-stream spatiotemporal prediction module; The physical constraint expression is: ,in, The distance between the target vehicle and the vehicle in front; This is the critical safety distance; This is the Gamma distribution function.
9. The service area dynamic parking guidance method based on real-time trajectory perception according to claim 8, characterized in that: The information input section of the dual-stream spatiotemporal prediction module is a parking space status perception layer, which includes an vacant parking space topology map. and dynamic occupancy matrix The spatial coordinates of the parking space are defined as follows: ,in, East-west coordinates, North-south coordinates Indicates the orientation of the parking space.
10. The service area dynamic parking guidance method based on real-time trajectory perception according to claim 4, characterized in that: Regarding the The algorithm is improved by introducing a dynamic adjustment factor into the heuristic function, expressed as follows: ,in, From the starting point to The actual cost, for Estimated cost to the destination, dynamic adjustment factor It should meet the following requirements: .