Intelligent parking guidance and reverse vehicle searching system based on multi-source heterogeneous data fusion

Abstract

The invention relates to the technical field of machine learning, and particularly discloses an intelligent parking guidance and reverse vehicle searching system based on multi-source heterogeneous data fusion. The system comprises a data perception fusion layer, a dynamic prediction decision-making layer and a user service interaction layer, realizes multi-step advanced probability prediction and dynamic optimal path planning of a parking space state through fusion of geomagnetic detection, video identification, payment flow, traffic situation and activity information, and realizes the optimal path planning of the parking space state based on multi-mode induction and live-action AR reverse vehicle searching guidance. And the parking efficiency and the user experience are improved.

Description

Technical Field

[0001] This invention belongs to the field of machine learning technology, specifically relating to a smart parking guidance and reverse vehicle search system that integrates multi-source heterogeneous data. Background Technology

[0002] In the field of smart cities and intelligent transportation systems, the optimized allocation and efficient utilization of parking resources are key to alleviating urban traffic congestion and improving citizens' travel experience. Smart parking systems, as an important component of this, aim to solve the parking problem through information technology and intelligent methods.

[0003] Among them, parking guidance and reverse vehicle search technology are core functional modules of the smart parking system. Their goal is to guide drivers to quickly find available parking spaces in real time and efficiently locate their vehicles when leaving. This technology typically relies on multiple sensors and data sources to perceive parking space status and plan routes.

[0004] Existing technologies primarily rely on current parking space availability data for guidance, but they cannot predict changes in parking space occupancy in the near future, leading to situations where parking spaces are already occupied when drivers arrive, thus creating a false sense of vacancy.

[0005] Meanwhile, although the existing system has access to various heterogeneous data sources such as geomagnetic data, camera data, and payment records, it has failed to deeply integrate the spatiotemporal correlation of information such as vehicle entry and exit records, surrounding road congestion status, and large-scale events in neighboring areas, which limits the accuracy of the prediction model.

[0006] These problems are particularly prominent during peak hours or around large events, severely reducing the accuracy of parking guidance and user trust. Therefore, there is an urgent need for a smart parking guidance and reverse vehicle search system that can achieve deep integration of multi-source heterogeneous data. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a smart parking guidance and reverse vehicle search system based on the fusion of multi-source heterogeneous data, which aims to overcome the problems of insufficient accuracy of parking guidance and low user trust caused by the inability to predict the future occupancy status of parking spaces and the failure to deeply integrate the spatiotemporal correlation of multi-source heterogeneous data in the existing technology.

[0008] This invention provides a smart parking guidance and reverse vehicle search system based on multi-source heterogeneous data fusion, comprising a data perception and fusion layer, a dynamic prediction and decision-making layer, and a user service interaction layer. The data perception and fusion layer is responsible for collecting and preprocessing various heterogeneous data from within the parking lot and the urban traffic environment. The dynamic prediction and decision-making layer, based on the fusion features output by the data perception and fusion layer, performs multi-step forward prediction of parking space status and dynamic planning of the optimal route. The user service interaction layer is responsible for converting the decision results into specific guidance instructions and vehicle search information, and interacting with users through terminal devices.

[0009] The data perception and fusion layer specifically includes a geomagnetic parking space detection module, a video stream analysis module, a payment transaction analysis module, an urban traffic situation access module, and a large-scale event information subscription module.

[0010] The geomagnetic parking space detection module continuously collects magnetic field change data of each parking space at a preset sampling frequency, and determines the instantaneous occupancy status of the parking space through the built-in signal processing algorithm.

[0011] The video stream analysis module is deployed at key access nodes in the parking lot to perform real-time license plate recognition and motion trajectory tracking of vehicles entering and exiting. Its output data is time-stamped and spatially correlated with the geomagnetic detection data.

[0012] The payment transaction analysis module obtains real-time vehicle exit payment records from the parking lot's central payment system and uses this information to infer the release time of the corresponding parking space. The urban traffic situation access module obtains real-time average vehicle speed and congestion index of roads surrounding the parking lot from the urban traffic command center via a standard data interface.

[0013] The large-scale event information subscription module proactively retrieves the schedules of registered large-scale events and their expected pedestrian traffic data for the next few hours from the city's public information platform in the vicinity of the parking lot.

[0014] All the raw data collected by the above modules are sent to a unified multi-source data spatiotemporal alignment and feature extraction engine. This engine first unifies all data streams to the same spatiotemporal coordinate system, and then extracts deep fusion feature vectors that can characterize the evolution of parking space status, vehicle entry and exit behavior patterns, and external environmental disturbance factors.

[0015] The dynamic prediction decision layer receives a deep fusion feature vector from the data perception fusion layer, and integrates a parking space status prediction unit and a path planning optimization unit.

[0016] The parking space status prediction unit adopts a multivariate time series prediction model based on an attention mechanism. The model takes historical parking space occupancy sequence, real-time vehicle entry and exit records, surrounding road congestion index, and future large-scale event information as joint inputs. Through the multi-layer encoder structure inside the model, it learns the complex nonlinear mapping relationship between the above multi-source variables and the future occupancy status of parking spaces, and then outputs the probability prediction value of each parking space status within the next 5 to 30 minutes time scale.

[0017] The path planning optimization unit constructs a dynamic path planning model with the goal of minimizing the user's total search time, based on the output of the parking space status prediction unit, the real-time traffic conditions of each lane in the current parking lot, and the physical size constraints of the user's vehicle.

[0018] The model abstracts the internal road network of a parking lot into a weighted directed graph. The weights in the graph are dynamically adjusted based on the predicted parking space occupancy probability and the real-time congestion of the passage. Then, an improved Dijkstra algorithm is used to solve the optimal driving path from the parking lot entrance to the target vacant parking space.

[0019] The user service interaction layer includes a parking guidance module and a reverse vehicle search module. The parking guidance module receives optimized route information from the dynamic prediction and decision-making layer and converts it into specific graphical guidance interfaces and voice prompts, which are then disseminated multimodally through off-site guidance screens, on-site guidance indicator lights, and user mobile terminal applications.

[0020] The reverse vehicle search module is activated when the user leaves. It retrieves the vehicle parking location image recorded and associated by the video stream analysis module when the user entered the parking lot, and combines it with the high-precision Bluetooth beacon or ultra-wideband positioning network inside the parking lot to provide the user with the optimal walking route planning and real-scene AR navigation guidance from the current location to their parked vehicle.

[0021] As a preferred embodiment of the present invention, the multi-source data spatiotemporal alignment and feature extraction engine performs the following processing flow.

[0022] First, the engine assigns a uniform timestamp and spatial location label to all input data streams.

[0023] Secondly, the engine uses a spatiotemporal kriging interpolation method to repair missing or abnormal data points.

[0024] Finally, the engine uses a deep feature extraction network to automatically learn a fusion feature representation from the repaired multi-source data that can comprehensively reflect the short-term fluctuations and long-term trends of parking space status.

[0025] Furthermore, the training process of the attention-based multivariate time series prediction model is as follows.

[0026] The model uses multi-source heterogeneous data from multiple consecutive time slices in historical data as training samples, takes data from the previous time to the current time as input, and takes the actual parking space status from the next time to multiple future time as prediction target. It minimizes the cross-entropy loss function between the prediction target and the model output, and uses the backpropagation algorithm to iteratively optimize all parameters inside the model until the prediction accuracy of the model on the independent validation set reaches a preset threshold.

[0027] Furthermore, the improved Dijkstra algorithm used in the path planning optimization unit has undergone the following adaptive modifications. When calculating the path cost, the algorithm not only considers the physical length of the passage but also introduces a dynamic adjustment factor, which is a weighted sum of the probability of the target parking space being occupied in the future and the current traffic resistance of the passage.

[0028] The algorithm prioritizes exploring path branches that lead to parking spaces with low predicted occupancy probability and low passage resistance.

[0029] Furthermore, the real-view AR navigation guidance function of the reverse vehicle search module is achieved through the following technical steps. The module first captures a real-time environmental video stream using the user's smartphone's built-in camera.

[0030] The module then uses visual SLAM technology to process the video stream and estimate the user's position and orientation relative to the fixed beacon in the parking lot in real time.

[0031] Finally, the module overlays and renders the pre-calculated walking path and directional arrows onto the real-time video feed, creating an intuitive augmented reality navigation interface.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] 1. This invention achieves deep spatiotemporal alignment and feature fusion of multi-source heterogeneous data such as geomagnetic detection, video recognition, payment records, and urban traffic and activity information by constructing a data perception fusion layer, providing a solid data foundation for accurate prediction.

[0034] 2. The dynamic prediction decision layer introduces a multivariate time series prediction model based on the attention mechanism, which can capture the complex relationship between parking space status and various internal and external factors, and realize multi-step advance probability prediction of the future occupancy status of parking spaces, fundamentally avoiding false vacancy misleading.

[0035] 3. The route planning and optimization unit combines dynamic prediction results with real-time channel conditions to calculate and optimize the route cost in real time, ensuring that users are provided with the shortest total search time for the travel route.

[0036] 4. The multimodal guidance and real-scene AR reverse car-finding guidance provided by the user service interaction layer greatly improve the smoothness and ease of operation of the user experience throughout the entire parking and car-finding process.

[0037] 5. The entire system improves the accuracy, timeliness, and user satisfaction of parking guidance through a progressive technical solution, and is especially suitable for complex parking scenarios such as peak hours and areas around large events. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the overall technical solution architecture of the intelligent parking guidance and reverse vehicle search system based on multi-source heterogeneous data fusion proposed in this invention.

[0039] Figure 2 This is a schematic diagram of the core principle framework of the multi-source data spatiotemporal alignment and feature extraction engine in this invention;

[0040] Figure 3 This is a schematic diagram of the core principle framework of the multivariate time series prediction model based on the attention mechanism in this invention;

[0041] Figure 4 This is a flowchart of the dynamic path planning model of the path planning optimization unit in this invention.

[0042] Figure 5 This is a schematic diagram illustrating the interaction relationship between the multimodal guidance and reverse vehicle search functions in the user service interaction layer of this invention. Detailed Implementation

[0043] This embodiment details the technical implementation of a smart parking guidance and reverse vehicle search system based on multi-source heterogeneous data fusion. Please refer to the appendix. Figure 1 The overall architecture of the system is divided into three main layers: the data perception and fusion layer, the dynamic prediction and decision-making layer, and the user service interaction layer.

[0044] The data perception and fusion layer, as the system's foundational data source, is responsible for collecting various types of data from within the parking lot and the urban traffic environment and performing preliminary processing. The dynamic prediction and decision-making layer is the system's core computing engine, using the fusion features provided by the data perception and fusion layer to predict the future status of parking spaces and optimize driving routes.

[0045] The user service interaction layer is the interface through which the system interacts directly with end users. It is responsible for transforming decision results into specific guidance information and presenting it to users through various terminal devices.

[0046] The data perception and fusion layer includes five key data acquisition modules: geomagnetic parking space detection module, video stream analysis module, payment transaction analysis module, urban traffic situation access module, and large-scale event information subscription module.

[0047] The specific implementation of the geomagnetic parking space detection module is as follows. The module buries a geomagnetic sensor node beneath the ground of each parking space. This node continuously monitors the vertical intensity changes of the Earth's magnetic field at a sampling frequency of 10 times per second.

[0048] When the local magnetic sensor detects a sudden change in magnetic field strength greater than a preset threshold of 50 nT, the microprocessor built into the module will start a digital filtering algorithm to reduce noise in the original signal. Then, a classifier based on support vector machine will be used to analyze the feature vector of the filtered signal in real time to determine whether the current parking space is occupied or vacant with at least 95% confidence.

[0049] Each geomagnetic sensor node packages and sends the detection results and corresponding timestamps to the area gateway inside the parking lot via a Bluetooth Low Energy 5.0 mesh network.

[0050] The specific deployment and workflow of the video stream analysis module are as follows: One high-definition network camera is deployed at every 50m interval at all entrance and exit lanes of the parking lot, as well as at major intersections of driving lanes.

[0051] These cameras continuously capture video stream data at a rate of 25fps. The video stream is transmitted in real time to an edge computing server deployed in the parking lot's server room.

[0052] The video analysis software running on the server first uses the YOLOv5 object detection algorithm to locate all vehicles in each frame of the image, and then uses the DeepSORT algorithm to track each detected vehicle across frames to generate a complete vehicle motion trajectory.

[0053] For each vehicle that enters the tracking state, the system will capture an image of the license plate area from the frame with the clearest motion trajectory and call the OCR optical character recognition engine to identify the complete license plate number.

[0054] Finally, the system associates the start and end coordinates of the vehicle's trajectory, the timestamp, and the identified license plate number to form a complete vehicle passage record.

[0055] The payment transaction record parsing module uses the RESTful API provided by the parking lot central payment system to poll for the latest vehicle exit payment records every 2 seconds. Each payment record includes fields such as license plate number, exit timestamp, payment amount, and payment method.

[0056] This module maintains a database mapping license plates to parking space numbers. This database is dynamically updated by the video stream analysis module when a vehicle enters the parking lot by recording the final parking space number where the vehicle stops.

[0057] When an exit payment record is parsed, the module immediately queries the mapping database to find the parking space number corresponding to the license plate, and calculates the approximate release time of the parking space by combining the exit timestamp.

[0058] Considering the potential delay between the vehicle actually leaving and the payment being completed, the module introduces a 1-minute time buffer correction, marking the parking space release time as the exit timestamp minus 1 minute.

[0059] The urban traffic situation access module establishes a secure connection with the data open platform of the urban traffic command center via the HTTPS protocol. The module sends a data request to the platform every 30 seconds. The request message contains the geographical location coordinates (latitude and longitude) of the parking lot and a circular area with a radius of 1km.

[0060] The response data returned by the platform includes the real-time average vehicle speed of all main roads in the area, in km / h, and a comprehensive congestion index, which is a decimal between 0 and 1, where 0 indicates completely smooth traffic and 1 indicates severe congestion.

[0061] The module parses and caches this data to provide input for subsequent predictive decisions.

[0062] The large-scale event information subscription module maintains a persistent connection with the city's public information platform's WebSocket service. Upon startup, the module subscribes to information on all registered large-scale events within a 3km radius of the parking lot.

[0063] When the platform releases new event announcements or updates existing event information, it will actively push them to this module via a WebSocket connection.

[0064] The data packets pushed are in JSON format and contain key fields such as event name, event start time, event end time, event location latitude and longitude, and expected visitor volume level 1 to 5.

[0065] The module will take these activity information into consideration 2 hours in advance.

[0066] Please refer to the attached document. Figure 2 All the data generated by the above five modules are ultimately integrated into a unified multi-source data spatiotemporal alignment and feature extraction engine.

[0067] The engine first establishes a local Cartesian coordinate system with the parking lot entrance as the origin, and then transforms all data with spatial location information, such as parking space locations detected by geomagnetic detection, vehicle coordinates tracked by video, and latitude and longitude of activity locations, into this unified coordinate system.

[0068] For time alignment, the engine uses a high-precision Network Time Protocol (NTP) server to timestamp all input data with a uniform time accuracy down to the millisecond level. In cases of missing data, such as a geomagnetic sensor losing 2 minutes of data due to communication interruption, the engine will initiate a repair algorithm based on spatiotemporal kriging interpolation.

[0069] The algorithm takes all available similar data within a 50m radius around the missing point and within 5 minutes before and after the time as input, and calculates an optimal interpolation estimate to fill the data gap.

[0070] After completing spatiotemporal alignment and missing value repair, the engine activates its internal deep feature extraction network. This network is a fully connected neural network with three hidden layers.

[0071] The number of input layer nodes is consistent with the total number of feature dimensions of multi-source data, such as 5-dimensional geomagnetic signal features, 8-dimensional video trajectory features, 2-dimensional traffic index, and 3-dimensional activity information.

[0072] The first hidden layer contains 128 neurons and uses the ReLU activation function; the second hidden layer contains 64 neurons; the third hidden layer, which is the output layer, contains 32 neurons and outputs a 32-dimensional deep fusion feature vector.

[0073] This feature vector comprehensively represents the short-term fluctuation characteristics of parking space status, such as the variance of magnetic field changes in the last minute, as well as long-term trends, such as the frequency of the parking space being occupied in the past hour. It also incorporates the influence of the external environment, such as the potential impact of surrounding road congestion on vehicle arrival rates.

[0074] This feature vector is updated every 10 seconds and is sent to the dynamic prediction decision layer as the final output of the data perception fusion layer.

[0075] The dynamic prediction decision layer comprises two core units: a parking space status prediction unit and a route planning optimization unit. Please refer to the appendix. Figure 3 The parking space status prediction unit adopts a multivariate time series prediction model based on an attention mechanism.

[0076] The input to the model is a multi-source data sequence within a time window, specifically including the following data from the past 10 time steps, with each time step spaced 30 seconds apart: the occupancy status sequence of all parking spaces, the vehicle entry and exit record sequence output by the video stream analysis module, the average vehicle speed sequence of surrounding roads provided by the urban traffic situation access module, and the future event pedestrian flow index provided by the large-scale event information subscription module.

[0077] The model first processes these input sequences through an encoder section.

[0078] The encoder consists of two layers of bidirectional long short-term memory networks, with 64 hidden units in each layer.

[0079] Bidirectional long short-term memory networks process input sequences from both the forward and reverse directions to capture temporal dependencies.

[0080] The hidden state output by the encoder is then fed into an attention layer. The mechanism of the attention layer is to calculate the importance weight of each input time step for predicting the future parking space state. Specifically, for the future time step to be predicted... At each time step, the model calculates the attention weights. as follows:

[0081] ;

[0082] in The encoder is in the first The hidden state at each time step. Is the decoder predicting the first... The hidden state at each time step It is a scoring function based on dot product. Attention weights. Represents the first in history Information at any given moment is useful for predicting the future. The importance of the status of parking spaces at all times.

[0083] After all weights have been calculated, the model sums the hidden states of the encoder at all time steps according to their attention weights to obtain a context vector. This context vector integrates the most relevant information from the historical sequence, and is then concatenated with the hidden states of the decoder. Together, they pass through a fully connected layer and a softmax activation function to output the predicted probability value of each parking space being vacant at six time points: 5 min, 10 min, 15 min, 20 min, 25 min, and 30 min. This probability value is a decimal between 0 and 1.

[0084] The route planning optimization unit receives the output from the parking space status prediction unit, which is the probability of each parking space being vacant at multiple future time points. Please refer to the appendix. Figure 4 This unit first constructs a weighted directed graph describing the internal road network of the parking lot. Each node in the graph represents a key location, including the parking lot entrance, exit, lane intersections, and the parking spot in front of each parking space. Each directed edge in the graph represents a passable lane, and the initial weight of the edge is the physical length of the lane, in meters.

[0085] To dynamically reflect the uncertainty of traffic conditions and parking space availability, the route planning optimization unit adjusts the weight of each edge in the graph in real time. The adjustment factor consists of two parts. The first part is the average occupancy probability of the target parking space over a future period of time.

[0086] For example, for a path leading to parking space A, the unit calculates the average probability that parking space A will be occupied within 5 minutes after the estimated time when a vehicle arrives at parking space A (based on path length and average vehicle speed). The second part is the current traffic resistance of the passage, which is quantified by the number of vehicles passing through the passage per unit time, as statistically analyzed by the video stream analysis module.

[0087] Traffic resistance The calculation formula is: , where N is the number of vehicles that have passed through the channel in the last minute.

[0088] Final dynamic weights for:

[0089] ;

[0090] in It is the static physical length. R is the predicted occupancy probability of the target parking space, and R is the passageway resistance. and These are preset weighting coefficients, with typical values ​​of 0.5 and 0.3 respectively.

[0091] After the weights are dynamically updated, the path planning optimization unit uses an improved Dijkstra's algorithm to find the optimal path from the parking lot entrance to all potential vacant parking spaces. The standard Dijkstra's algorithm iteratively explores the shortest path from the starting point to all other nodes.

[0092] The improvement in this unit is that when the algorithm explores a node, it not only considers the cumulative path cost from the starting point to the node, but also proactively evaluates the cost from the node to its downstream target parking space. This downstream cost includes the dynamic prediction weight of the target parking space.

[0093] The algorithm maintains a priority queue, where the priority of a node is composed of two parts: the actual cost from the starting point to that node. and a heuristic cost , It is set as the estimated minimum dynamic cost from this node to the target parking space.

[0094] The algorithm prioritizes expanding those The node with the smallest value is selected to more efficiently guide drivers to parking spaces with low predicted occupancy probability and unobstructed passageways.

[0095] Finally, the algorithm outputs a driving path with the minimum total cost from the entrance to the specified target parking space, which is represented as a sequence of nodes.

[0096] The user service interaction layer includes a parking guidance module and a reverse vehicle search module. Please refer to the appendix. Figure 5 The specific workflow of the parking guidance module is as follows.

[0097] The module receives the optimal path information from the path planning and optimization unit.

[0098] For off-site guidance, the module sends summary information of the optimal route, such as the target area and estimated search time, to a large full-color LED guidance screen set up near the parking lot entrance via a 4G / 5G network.

[0099] The guidance screen graphically displays the predicted number of remaining parking spaces and recommended driving directions for each area inside the parking lot.

[0100] For in-park guidance, the module controls height guidance indicator lights installed above the lanes. These lights use tri-color LEDs: green indicates a high probability of available parking spaces ahead, yellow indicates limited spaces, and red indicates no available spaces. For users' mobile devices, the module pushes detailed route planning results to the user's smartphone application.

[0101] The application interface displays a map of the parking lot, clearly marking the optimized driving route from the user's current location to the recommended parking space, supplemented by text prompts and voice announcements, such as "Turn left 50m ahead, the target parking space is on the right."

[0102] The reverse vehicle search module is activated when the user finishes shopping or is ready to leave.

[0103] Users can activate this feature by clicking the car-finding button on the terminal application or scanning the QR code in the parking lot.

[0104] The module first obtains the user's current identity identifier through the application, which is usually the license plate number recorded upon entry.

[0105] Subsequently, the module queries the database of the video stream analysis module to retrieve the parking space number where the vehicle finally stopped when it entered the parking lot, and retrieves the vehicle parking feature image captured by the camera near the parking space.

[0106] Meanwhile, the module scans for pre-deployed iBeacon beacon signals in the field using the Bluetooth 5.1 chip built into the smartphone.

[0107] Each beacon broadcasts its unique identifier and signal strength indicator value (RSSI) 10 times per second.

[0108] The module uses a triangulation algorithm to calculate the high-precision coordinates of the user's current location based on the RSSI values ​​of at least three different beacons, with a positioning accuracy of less than 1 meter.

[0109] After obtaining the user's current location and the vehicle's location, the reverse vehicle search module activates its augmented reality navigation function.

[0110] The module uses the smartphone's rear camera to start capturing a real-time video stream of the environment.

[0111] The video stream is fed into a processing engine based on visual SLAM technology.

[0112] The engine extracts ORB feature points from the video sequence and tracks the motion of these feature points between consecutive frames. At the same time, it combines the data from the phone's built-in inertial measurement unit (IMU) to estimate the 6-DOF pose of the phone's camera in the parking lot coordinate system in real time.

[0113] At the same time, the module calculates an optimal walking route based on the user's current location and the vehicle's parking location, avoiding densely populated areas and obstacles.

[0114] Finally, the module uses the OpenGLES graphics library to accurately overlay and render the calculated path lines, direction arrows, and distance prompts onto the real-time camera preview.

[0115] The directional arrows adjust their direction in real time based on the user's movement and the orientation of their phone, ensuring the user always moves in the correct direction until they find their vehicle. The entire AR navigation interface is intuitive and clear, greatly simplifying the process of finding your car in complex parking lot environments.

[0116] The system described in this embodiment achieves deep integration of multi-source heterogeneous data through a data perception fusion layer, providing high-quality input for prediction.

[0117] The dynamic prediction decision layer utilizes advanced prediction models and optimization algorithms to ensure the accuracy and efficiency of parking space status prediction and route planning.

[0118] The user service interaction layer seamlessly transmits the system's intelligent decision-making results to users through diverse interaction methods, ultimately building an efficient, accurate, and user-friendly intelligent parking guidance and reverse vehicle search system.

[0119] This system is particularly suitable for scenarios with high parking demand and complex traffic conditions, such as commercial centers, transportation hubs, and areas around large event venues. It can alleviate parking difficulties and improve parking lot operation efficiency and user satisfaction.

[0120] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0121] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart parking guidance and reverse vehicle search system that integrates multi-source heterogeneous data, characterized in that, include: The data perception and fusion layer is used to collect and preprocess various heterogeneous data from inside the parking lot and the urban traffic environment. The data perception and fusion layer includes a geomagnetic parking space detection module, a video stream analysis module, a payment transaction analysis module, an urban traffic situation access module, and a large-scale event information subscription module. The dynamic prediction decision layer is used to perform multi-step advance prediction of parking space status and dynamic planning of the optimal path based on the fusion features output by the data perception fusion layer. The dynamic prediction decision layer includes a parking space status prediction unit and a path planning optimization unit. The parking space status prediction unit adopts a multivariate time series prediction model based on attention mechanism. It uses historical parking space occupancy sequence, real-time vehicle entry and exit records, surrounding road congestion index and future large-scale event information as joint inputs. Through the multi-layer encoder structure inside the model, it learns the complex nonlinear mapping relationship between multi-source variables and future parking space occupancy status, and outputs the probability prediction value of each parking space status in the future time scale. The path planning optimization unit constructs a dynamic path planning model with the goal of minimizing the total search time of users based on the output of the parking space status prediction unit, the real-time traffic conditions of each lane in the current parking lot, and the physical size constraints of the user's vehicle. The model abstracts the internal road network of the parking lot into a weighted directed graph. The weights in the graph are dynamically adjusted according to the predicted parking space occupancy probability and the real-time lane congestion. An improved Dijkstra algorithm is used to solve the optimal driving path from the parking lot entrance to the target vacant parking space. The user service interaction layer is used to transform decision results into specific guidance instructions and vehicle location guidance information, and interact with users through terminal devices. The user service interaction layer includes a parking guidance module and a reverse vehicle location module.

2. The intelligent parking guidance and reverse vehicle search system based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The parking guidance module receives optimized route information from the dynamic prediction decision layer and converts it into a graphical guidance interface and voice prompts, which are then disseminated in a multimodal manner through off-site guidance screens, on-site guidance indicator lights, and user mobile terminal applications. The reverse vehicle search module is activated when the user leaves. By retrieving the vehicle parking location image recorded and associated by the video stream analysis module when the user entered, and combining it with the high-precision Bluetooth beacon or ultra-wideband positioning network inside the parking lot, it provides the user with the optimal walking route planning and real-scene AR navigation guidance from the current location to their parked vehicle.

3. The intelligent parking guidance and reverse vehicle search system based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The data perception fusion layer also includes a multi-source data spatiotemporal alignment and feature extraction engine. The engine unifies all data streams to the same spatiotemporal coordinate system and extracts deep fusion feature vectors that can characterize the evolution of parking space status, vehicle entry and exit behavior patterns, and external environmental disturbance factors.

4. The intelligent parking guidance and reverse vehicle search system based on multi-source heterogeneous data fusion according to claim 3, characterized in that, The multi-source data spatiotemporal alignment and feature extraction engine performs the following processing steps: assigning a unified timestamp and spatial location label to all input data streams; repairing missing or abnormal data points using a spatiotemporal kriging interpolation method; and automatically learning a fusion feature representation from the repaired multi-source data using a deep feature extraction network that can comprehensively reflect the short-term fluctuations and long-term trends of parking space status.

5. The intelligent parking guidance and reverse vehicle search system based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The training process of the attention-based multivariate time series prediction model is as follows: The model uses multi-source heterogeneous data from multiple consecutive time slices in historical data as training samples, takes the data from the previous time to the current time as input, and takes the actual parking space status from the next time to multiple future times as the prediction target. By minimizing the cross-entropy loss function between the prediction target and the model output, and using the backpropagation algorithm to iteratively optimize all parameters inside the model, the prediction accuracy of the model on the independent validation set reaches a preset threshold.

6. The intelligent parking guidance and reverse vehicle search system based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The improved Dijkstra algorithm has been adapted as follows: when calculating the path cost, the algorithm not only considers the physical length of the channel, but also introduces a dynamic adjustment factor, which is a weighted sum of the occupancy probability of the target parking space in the future and the current traffic resistance of the channel. The algorithm prioritizes exploring path branches that lead to parking spaces with low predicted occupancy probability and low passage resistance.

7. The intelligent parking guidance and reverse vehicle search system based on multi-source heterogeneous data fusion according to claim 2, characterized in that, The real-scene AR navigation guidance function of the reverse vehicle search module is realized through the following technical steps: capturing real-time environmental video streams through the built-in camera of the user's smartphone; processing the video streams using visual SLAM technology to estimate the user's position and attitude relative to the fixed beacon in the parking lot in real time; and overlaying and rendering the pre-calculated walking path and directional arrows onto the real-time video screen to form an intuitive augmented reality navigation interface.

8. The intelligent parking guidance and reverse vehicle search system based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The geomagnetic parking space detection module continuously collects magnetic field change data of each parking space at a preset sampling frequency, and determines the instantaneous occupancy status of the parking space through the built-in signal processing algorithm. The video stream analysis module is deployed at key access nodes of the parking lot to perform real-time license plate recognition and motion trajectory tracking of vehicles entering and exiting; The payment transaction analysis module obtains real-time vehicle exit payment records from the central payment system of the parking lot and uses the license plate information to infer the release time of the corresponding parking space. The urban traffic situation access module obtains the real-time average vehicle speed and congestion index of the roads surrounding the parking lot from the urban traffic command center through a standard data interface. The large-scale event information subscription module actively retrieves the schedules of registered large-scale events and their expected pedestrian traffic data for the next few hours from the city's public information platform in the vicinity of the parking lot.

9. The intelligent parking guidance and reverse vehicle search system based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The attention mechanism calculation process in the parking space state prediction unit is as follows: For the k-th time step to be predicted, the model calculates the attention weight for each input time step. The attention weight is calculated by the dot product score function of the encoder's hidden state at each time step and the decoder's hidden state at the previous moment of prediction. The context vector is obtained by summing the hidden states of all time steps according to their attention weights. The context vector is concatenated with the hidden state of the decoder and then passed through a fully connected layer and a softmax activation function to output the predicted probability values ​​of each parking space being vacant at multiple future time points.

10. A smart parking guidance and reverse vehicle search system based on multi-source heterogeneous data fusion according to claim 2, characterized in that, The specific workflow of the video stream analysis module includes: using the YOLOv5 target detection algorithm to locate all vehicles in each frame of the image; and using the DeepSORT algorithm to perform cross-frame tracking on each detected vehicle to generate a complete vehicle motion trajectory. The image of the vehicle license plate area is captured from the frame with the clearest vehicle trajectory, and the OCR optical character recognition engine is used to recognize the complete license plate number. The starting and ending coordinates of the vehicle trajectory, the timestamp, and the identified license plate number are linked to form a complete vehicle passage record.

Images