A parking lot exit vehicle congestion prediction and scheduling method, medium and system

Abstract

The present invention provides a method, medium and system for predicting and dispatching vehicle congestion at a parking lot exit, which belongs to the technical field of commercial data supervision and processing. By constructing a tree-structured neural network composed of multiple lightweight neural networks, vehicle traffic data is first collected at the parking lot exit, and data expansion is performed using a generative adversarial network. The system uses a memory allocation optimization algorithm to optimize the network memory, and divides the time series data into training samples containing vehicle travel time, quantity and type. The neural network extracts features layer by layer, successively obtains feature vectors and multiple prediction features, establishes a vehicle congestion prediction model, and generates a vehicle diversion scheduling plan based on this. Finally, the system executes the scheduling plan and collects the execution results, continuously updates the training data set, and realizes closed-loop optimization of intelligent traffic flow control. The method, medium and system of the present invention solve the problem that the prior art is difficult to predict vehicle congestion and thus conduct active scheduling and prevention.

Description

Technical Field

[0001] The present invention belongs to the technical field of commercial data monitoring and processing, and in particular, relates to a method, medium and system for predicting and dispatching vehicle congestion at a parking lot exit. Background Art

[0002] With the acceleration of urbanization and the vigorous development of modern business society, the scale of parking lots in large commercial complexes, office buildings and residential areas has continued to expand. Especially in the urban commercial center area, large parking lots with thousands of parking spaces have become the norm. This large-scale parking lot management has brought huge challenges to vehicle exit passage. Under the traditional parking lot management model, vehicles often get congested when passing through the exit, especially during peak hours in the morning and evening and during large-scale events. This problem is more prominent, which seriously affects the operation efficiency and user experience of the parking lot. At present, the industry mainly adopts a variety of technical solutions for parking lot exit congestion, including management methods based on fixed rules, responsive management methods based on real-time monitoring, management methods based on simple prediction models, and intelligent prediction methods based on single neural networks.

[0003] The management method based on fixed rules adjusts the number of exit channels through pre-set time periods, such as adding exit channels during expected peak hours. Although this method is simple to implement, it lacks flexibility and cannot cope with sudden changes in traffic volume. In addition, manual judgment often lags behind, resulting in poor control effects and difficulty meeting the needs of modern parking lot management. The responsive management method based on real-time monitoring is to install video surveillance and vehicle detectors in the exit area, count the number of queued vehicles in real time, and trigger the diversion mechanism when the preset threshold is reached. However, this method is a passive response and often takes measures only after congestion has formed. It is difficult to fundamentally solve the problem, and the system response time is long and the effect is poor.

[0004] Therefore, the current parking lot management field urgently needs a method that can predict vehicle congestion and conduct active scheduling and prevention to meet the intelligent management needs of modern large parking lots. Summary of the invention

[0005] In view of this, the present invention provides a method, medium and system for predicting and dispatching vehicle congestion at a parking lot exit, which can solve the problem that it is difficult to predict vehicle congestion and thus conduct active dispatch prevention in the prior art.

[0006] The present invention is achieved in that:

[0007] A first aspect of the present invention provides a method for predicting and dispatching vehicle congestion at a parking lot exit, comprising the following steps:

[0008] S10, constructing a tree-structured joint neural network, the joint neural network comprising four lightweight neural networks, a first lightweight neural network for extracting time features, a second lightweight neural network for extracting vehicle density features, a third lightweight neural network for extracting vehicle speed features, and a fourth lightweight neural network for feature fusion;

[0009] S20, collecting vehicle traffic data at the parking lot exit, and using a generative adversarial network to perform data expansion on the vehicle traffic data to obtain a training data set;

[0010] S30, optimizing the allocation of physical memory and graphics processor memory for the lightweight neural network based on a memory allocation optimization algorithm;

[0011] S40, dividing the training data set into a plurality of training samples according to the time series, each of the training samples comprising vehicle travel time, vehicle quantity, and vehicle type;

[0012] S50, inputting the training sample into the first lightweight neural network in the joint neural network to obtain a first eigenvector;

[0013] S60, inputting the first feature vector into other lightweight neural networks in the joint neural network in sequence to obtain multiple prediction features;

[0014] S70, establishing a vehicle congestion prediction model based on the multiple prediction features to obtain a vehicle congestion prediction result;

[0015] S80, generating a vehicle diversion scheduling plan according to the vehicle congestion prediction result;

[0016] S90: Execute the vehicle diversion scheduling plan, and collect execution results for updating the training data set.

[0017] On the basis of the above technical solution, the parking lot exit vehicle congestion prediction and dispatching method of the present invention can also be improved as follows:

[0018] Among them, the joint neural network includes four lightweight neural networks, the first lightweight neural network is used to extract time features, the second lightweight neural network is used to extract vehicle density features, the third lightweight neural network is used to extract vehicle speed features, and the fourth lightweight neural network is used for feature fusion. The outputs of the first lightweight neural network, the second lightweight neural network, and the third lightweight neural network are respectively connected to the input end of the fourth lightweight neural network.

[0019] Furthermore, the first lightweight neural network, the second lightweight neural network, and the third lightweight neural network all adopt a three-layer convolutional layer and a two-layer fully connected layer structure, and the fourth lightweight neural network adopts a four-layer fully connected layer structure.

[0020] Furthermore, the step of using a generative adversarial network for data expansion includes: establishing a generator network and a discriminator network, the generator network inputs a random noise vector and outputs simulated vehicle traffic data, the discriminator network compares the difference between the simulated vehicle traffic data and the real vehicle traffic data, and feeds the difference result back to the generator network for optimizing the generator network.

[0021] Furthermore, the memory allocation optimization algorithm includes a memory demand prediction equation, a memory allocation equation and a memory recovery equation. The memory demand prediction equation determines the memory demand based on the parameter quantity and calculation amount of each lightweight neural network. The memory allocation equation allocates physical memory and graphics processor memory based on the memory demand. The memory recovery equation releases memory based on the task completion status of the lightweight neural network.

[0022] Furthermore, the step of acquiring the training data set includes: collecting parking lot exit camera image data, extracting vehicle position information, vehicle speed information and vehicle type information in the image, collecting parking lot exit geomagnetic detector data to obtain vehicle travel time information, and combining the vehicle position information, vehicle speed information, vehicle type information and vehicle travel time information to form training data; the training data set is updated using a sliding time window method, retaining historical data for the past 90 days, and updating the training of the joint neural network every 24 hours.

[0023] Furthermore, the training process of the joint neural network adopts a staged training method. First, the first lightweight neural network, the second lightweight neural network and the third lightweight neural network are trained separately. After completing the single feature extraction training, the parameters of the three lightweight neural networks are fixed, and the fourth lightweight neural network is trained to complete feature fusion.

[0024] Furthermore, the vehicle congestion prediction model adopts a long short-term memory network structure, with the input being the multiple prediction features and the output being the probability of vehicle congestion within the next 30 minutes; the vehicle diversion scheduling plan includes an exit channel quantity adjustment plan and a vehicle diversion recommendation plan, the exit channel quantity adjustment plan is used to dynamically adjust the number of open exit channels, and the vehicle diversion recommendation plan is used to push recommended exit selection information to vehicle drivers.

[0025] The equations involved in the calculation process and their principles are explained in detail below.

[0026] The generation loss function of the generative adversarial network is specifically expressed as follows:

[0027] ;

[0028] In the formula, is the generator loss; Output the result for the generator; is a random noise vector, which follows a standard normal distribution ; It is the real traffic data; is the discriminator's judgment result on the generated data; is the smoothing loss term; is the weight coefficient and satisfies .

[0029] The memory requirement prediction equation is specifically expressed as follows:

[0030] ;

[0031] In the formula, For the Memory requirements of a lightweight neural network; is the parameter quantity; is the amount of calculation; is the batch size; is the hidden layer size; is the weight coefficient; is the prediction error term.

[0032] The memory allocation equation is specifically expressed as follows:

[0033] ;

[0034] In the formula, For the The amount of memory allocated for each lightweight neural network; is the task priority coefficient; is the total number of lightweight neural networks; The total amount of memory available on the system.

[0035] The time feature extraction matrix of the first lightweight neural network is specifically expressed as follows:

[0036] ;

[0037] In the formula, is the time feature matrix; For the The time segment eigenvalues; is the number of time segments; is the feature dimension.

[0038] The vehicle density feature vector of the second lightweight neural network is specifically expressed as follows:

[0039] ;

[0040] In the formula, is the vehicle density feature vector; is the vector of the number of small vehicles; is the number vector of medium-sized vehicles; is the large vehicle quantity vector; is the weight coefficient; is the density correction term.

[0041] The vehicle speed feature calculation of the third lightweight neural network is specifically expressed as follows:

[0042] ;

[0043] In the formula, is the vehicle speed characteristic; For the The instantaneous speed of the vehicle; is the acceleration; is the acceleration; is the measurement time interval; is the feature weight coefficient.

[0044] The feature fusion function of the fourth lightweight neural network is specifically expressed as follows:

[0045] ;

[0046] In the formula, For fusion features; As the activation function, the ReLU function is used; is the feature weight; is the bias term.

[0047] The prediction function of the vehicle congestion prediction model is specifically expressed as follows:

[0048] ;

[0049] In the formula, for The predicted value of congestion probability at the time; for The fusion characteristics of the moment; is the time step; is the prediction error term.

[0050] The vehicle diversion scheduling optimization objective function is specifically expressed as follows:

[0051] ;

[0052] In the formula, To optimize the goal; For the Average travel time per exit; For the The queue length of the lane; is the weight coefficient; is the target weight.

[0053] The construction principles and meanings of these equations are as follows:

[0054] 1. The generated loss function uses the L2 norm to measure the difference between the generated data and the real data, and introduces a smoothing term to improve the continuity of the generated data;

[0055] 2. The memory demand prediction equation takes into account the key parameter indicators of the neural network and adopts a linear weighted modeling method;

[0056] 3. The memory allocation equation is based on the principle of on-demand allocation, taking into account task priority and adopting a proportional allocation method;

[0057] 4. The time feature extraction matrix uses a two-dimensional structure to store multi-dimensional time series features;

[0058] 5. The vehicle density feature vector takes into account the differences in space occupation of different types of vehicles and introduces a density correction term;

[0059] 6. The vehicle speed characteristic calculation takes into account the combined effects of speed, acceleration and jerk, and uses exponential decay weights;

[0060] 7. The feature fusion function adopts a combination of weighted summation and nonlinear activation;

[0061] 8. The congestion prediction model uses LSTM structure to capture temporal dependencies;

[0062] 9. The scheduling optimization objective function balances the two key indicators of travel time and queue length.

[0063] The derivation and optimization process of each equation is explained in detail below.

[0064] 1. Generate loss function The derivation process:

[0065] Step 1: First consider the basic GAN loss, which only contains the discriminator feedback term ;

[0066] Step 2: In order to improve the authenticity of the generated data, the L2 distance term is introduced , used to constrain the difference between generated data and real data;

[0067] Step 3: Considering the continuity characteristics of traffic data, add a smoothing loss term , which is defined as:

[0068] ;

[0069] Step 4: By weight coefficient Balance the contributions of the three losses to form the final loss function.

[0070] The effect of this equation is:

[0071] 1) The authenticity of the generated data is guaranteed;

[0072] 2) Ensure the smooth continuity of time series data;

[0073] 3) Improved the diversity of generated data.

[0074] 2. Memory Demand Prediction Equation The derivation process:

[0075] Step 1: Analyze the sources of memory usage during neural network runtime, including parameter storage, calculation process, batch processing, and intermediate states;

[0076] Step 2: Establish a linear relationship for each memory usage factor:

[0077] (parameter storage);

[0078] (Calculate overhead);

[0079] (Batch processing overhead);

[0080] (hidden state overhead);

[0081] Step 3: Consider prediction error , combined to form the final prediction equation.

[0082] The effects of this equation are: 1) accurately predicting memory requirements; 2) supporting dynamic memory allocation; 3) optimizing memory usage efficiency.

[0083] 3. Memory allocation equation The derivation process:

[0084] Step 1: Based on total memory constraints ;

[0085] Step 2: Consider the impact of task priority and introduce priority coefficient ;

[0086] Step 3: Use the proportional allocation principle to ensure that the allocation amount is proportional to the demand and priority;

[0087] Step 4: Obtain the final distribution equation through normalization.

[0088] The effects of this equation are: 1) ensuring fairness in memory allocation; 2) supporting priority task scheduling; 3) avoiding waste of memory resources.

[0089] 4. Temporal feature extraction matrix The derivation process:

[0090] Step 1: Divide the time series into segments, each segment has a fixed time window length;

[0091] Step 2: Extract features for each time window, including:

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] Step 3: Construct a two-dimensional matrix to store multi-dimensional features.

[0097] The effects of this equation are: 1) capturing time series characteristics; 2) preserving data temporal relationships; 3) supporting multi-dimensional feature expression.

[0098] 5. Vehicle density feature vector Determination process:

[0099] Step 1: Count different types of vehicles:

[0100] ;

[0101] ;

[0102] ;

[0103] Step 2: Consider the impact of vehicle type on road occupancy and introduce weight coefficients;

[0104] Step 3: Add density correction Handle special situations.

[0105] The effects of this equation are: 1) accurately reflecting road occupancy; 2) distinguishing the impact of different vehicle models; 3) adapting to complex road conditions.

[0106] 6. Vehicle speed characteristics Determination process:

[0107] Step 1: Consider the speed based on vehicle kinematics analysis , acceleration and jerk ;

[0108] Step 2: Introduce the time decay factor , highlighting the importance of recent data;

[0109] Step 3: Get comprehensive features through weighted average.

[0110] The effects of this equation are: 1) comprehensively describing the vehicle motion state; 2) highlighting the impact of recent data; 3) reflecting the dynamic characteristics of traffic flow.

[0111] The present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the above-mentioned parking lot exit vehicle congestion prediction and scheduling method.

[0112] The present invention provides a parking lot exit vehicle congestion prediction and dispatching system, which includes the above-mentioned computer-readable storage medium.

[0113] Compared with the prior art, the method, medium and system for predicting and dispatching parking lot exit vehicle congestion provided by the present invention have the following beneficial effects:

[0114] In response to the above technical problems, the present invention proposes a method for predicting and dispatching vehicle congestion at a parking lot exit. The method uses advanced neural network technology and optimization algorithms to achieve comprehensive perception and accurate prediction of the exit vehicle traffic characteristics, and based on this, a dynamic diversion dispatching strategy is formulated. Compared with the prior art, the present invention has the following significant technical advantages:

[0115] 1. Enhanced data perception capability. The present invention adopts a multi-sensor fusion method to not only collect the vehicle travel time data at the exit, but also extracts rich feature information such as the number of vehicles and vehicle models. At the same time, the original data is expanded using a generative adversarial network, further enhancing the representativeness and diversity of the data.

[0116] 2. Improved prediction modeling accuracy. The present invention constructs a tree-structured joint neural network that can effectively extract multi-dimensional features such as time series, vehicle density and vehicle speed, and fuse them into comprehensive prediction factors. Compared with a single prediction model, this feature fusion mechanism greatly improves the accuracy of vehicle congestion prediction.

[0117] 3. Optimizes system operation efficiency. The present invention proposes a dynamic memory allocation algorithm based on the memory demand characteristics of neural networks. The algorithm can reasonably allocate physical memory and graphics processor memory according to the computing load of each neural network, avoiding the waste of resources and the occurrence of computing bottlenecks.

[0118] 4. An intelligent dispatching strategy has been developed. Based on the accurate congestion prediction results, the present invention has designed a comprehensive dispatching scheme that takes into account the dynamic adjustment of exit channels and the diversion and guidance of vehicles. This combination of active control and passive guidance can better alleviate exit congestion and improve overall traffic efficiency.

[0119] In general, the present invention has made significant progress in data perception, model building, system optimization and scheduling strategies, solving the problem that the existing technology is difficult to predict vehicle congestion and thus conduct active scheduling prevention. BRIEF DESCRIPTION OF THE DRAWINGS

[0120] Figure 1 A flow chart of the method provided by the present invention;

[0121] Figure 2 This is a diagram of the joint neural network architecture;

[0122] Figure 3 Optimize the schematic diagram for memory allocation;

[0123] Figure 4 It is the traffic flow and speed change diagram;

[0124] Figure 5 Heat map of jam probability prediction. DETAILED DESCRIPTION

[0125] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0126] like Figure 1 As shown, it is a flow chart of a method for predicting and dispatching vehicle congestion at a parking lot exit provided by the present invention. The method comprises the following steps:

[0127] S10, constructing a tree-structured joint neural network, the joint neural network comprising four lightweight neural networks, a first lightweight neural network for extracting time features, a second lightweight neural network for extracting vehicle density features, a third lightweight neural network for extracting vehicle speed features, and a fourth lightweight neural network for feature fusion;

[0128] S20, collecting vehicle traffic data at the parking lot exit, and using a generative adversarial network to perform data expansion on the vehicle traffic data to obtain a training data set;

[0129] S30, optimizing the allocation of physical memory and GPU memory for lightweight neural network based on memory allocation optimization algorithm;

[0130] S40, dividing the training data set into a plurality of training samples according to the time series, each training sample including vehicle travel time, vehicle quantity, and vehicle type;

[0131] S50, inputting the training sample into the first lightweight neural network in the joint neural network to obtain a first eigenvector;

[0132] S60, inputting the first feature vector into other lightweight neural networks in the joint neural network in sequence to obtain multiple prediction features;

[0133] S70, establishing a vehicle congestion prediction model based on multiple prediction features to obtain a vehicle congestion prediction result;

[0134] S80, generating a vehicle diversion scheduling plan according to the vehicle congestion prediction result;

[0135] S90: Execute the vehicle diversion scheduling plan, and collect the execution results for updating the training data set.

[0136] The specific implementation methods of the above steps are described in detail below:

[0137] The specific implementation method of step S10 is to construct a tree-structured joint neural network. First, the joint neural network consists of four lightweight neural networks. The first lightweight neural network is used to extract time features, and adopts a structure of three convolutional layers and two fully connected layers. The second lightweight neural network is used to extract vehicle density features, and also adopts a structure of three convolutional layers and two fully connected layers. The third lightweight neural network is used to extract vehicle speed features, and the structure is consistent with the first two networks. The fourth lightweight neural network is responsible for fusing these three features, and adopts a structure of four fully connected layers. These four lightweight neural networks are combined in a tree structure, in which the outputs of the first, second, and third networks are respectively connected to the input of the fourth network. Such a joint neural network structure can make full use of the interrelationships between different features and improve the final prediction performance.

[0138] The specific implementation method of step S20 is to use a generative adversarial network to perform data expansion on the collected vehicle traffic data. First, a generator network and a discriminator network are established. The generator network takes a random noise vector as input and outputs simulated vehicle traffic data. The discriminator network compares the differences between these simulated data and real vehicle traffic data, and feeds the difference results back to the generator network for optimizing the parameters of the generator network. Through this generative adversarial training method, the richness and diversity of training data can be greatly increased, effectively alleviating the problem of insufficient data.

[0139] The specific implementation method of step S30 is to use a memory allocation optimization algorithm to allocate physical memory and graphics processor memory to the aforementioned lightweight neural network. The algorithm includes three key parts: a memory demand prediction equation, a memory allocation equation, and a memory recovery equation. The memory demand prediction equation uses a linear weighted method to predict memory demand based on indicators such as the number of parameters, calculation amount, batch size, and hidden layer size of each lightweight neural network. The memory allocation equation uses the principle of proportional allocation based on these prediction results and taking into account the task priority to reasonably allocate the total memory to each lightweight neural network. When a neural network task is completed, the memory recovery equation will be responsible for releasing the corresponding memory resources. This optimization method based on memory demand prediction and dynamic allocation can effectively improve memory usage efficiency and ensure that the computing tasks of each neural network can be smoothly executed.

[0140] The specific implementation of step S40 is to divide the collected vehicle traffic data into multiple training samples. Each training sample contains three key features: vehicle traffic time, vehicle quantity and vehicle type. Among them, vehicle traffic time can be obtained through the parking lot exit geomagnetic detector; vehicle quantity and vehicle type can be obtained by analyzing the image data extracted by the exit surveillance camera. Combining these feature information to form training samples can provide rich input data for subsequent neural network training.

[0141] The specific implementation method of steps S50 and S60 is to input the above training samples into the aforementioned joint neural network for feature extraction and fusion. First, the training samples are input into the first lightweight neural network to extract time features. The time features are stored in a two-dimensional matrix structure, including multi-dimensional features such as mean, variance, peak value and period. Then, these time features are input into the second and third lightweight neural networks in turn to extract vehicle density features and vehicle speed features. The vehicle density feature takes into account the spatial occupancy differences of different types of vehicles and is comprehensively reflected by weighted summation. The vehicle speed feature combines kinematic indicators such as speed, acceleration and jerk, and introduces exponential decay weights to highlight the importance of recent data. Finally, these three features are input into the fourth lightweight neural network, and feature fusion is performed by weighted summation and ReLU activation to generate the final fused feature vector. Such a step-by-step feature extraction and fusion mechanism can fully explore the correlation between different features and improve the performance of the prediction model.

[0142] The specific implementation of step S70 is to establish a vehicle congestion prediction model based on the above fusion features. The prediction model adopts a long short-term memory network (LSTM) structure, with the input being the fusion features of the latest k time steps, and the output being the prediction result of the vehicle congestion probability within the next 30 minutes. The LSTM network can effectively capture the long-term and short-term dependencies of time series data, thereby making more accurate congestion predictions. At the same time, the 30-minute prediction window can also provide sufficient time margin for subsequent scheduling decisions.

[0143] The specific implementation method of step S80 is to generate a vehicle diversion scheduling plan based on the vehicle congestion prediction results. The scheduling plan includes two parts: an exit channel number adjustment plan and a vehicle diversion suggestion plan. The exit channel number adjustment plan dynamically monitors the current exit congestion situation and adjusts the number of open exit channels in a timely manner according to the prediction results to alleviate congestion. The vehicle diversion suggestion plan pushes the best exit selection suggestion to the vehicle drivers who are about to enter the exit to guide the reasonable diversion of vehicles. This diversion scheduling strategy that takes into account both active regulation and passive guidance can effectively alleviate the exit congestion problem.

[0144] The specific implementation method of step S90 is to execute the above-mentioned vehicle diversion scheduling scheme and collect the execution results for updating the training data set. During the execution process, the actual vehicle traffic conditions and diversion effects will be continuously monitored, and the feedback information will be recorded. Then, the new data will be added to the original training data set to form an updated data set. This sliding time window update mechanism can ensure that the training data set remains timely and representative, improve the self-learning ability of the joint neural network, and thus continuously optimize the performance of prediction and scheduling.

[0145] The present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the above-mentioned parking lot exit vehicle congestion prediction and scheduling method.

[0146] The present invention provides a parking lot exit vehicle congestion prediction and dispatching system, which includes the above-mentioned computer-readable storage medium.

[0147] Specifically, the principle of the present invention is: a vehicle congestion prediction model based on a joint neural network is proposed, and a comprehensive intelligent control system is constructed by combining dynamic memory optimization and intelligent diversion scheduling. The key to this solution is that through the in-depth mining and fusion of multi-dimensional features, the dynamic characteristics of exit congestion can be more accurately portrayed, laying a solid foundation for subsequent prediction and regulation.

[0148] First, the present invention constructs a tree-structured joint neural network. The network consists of four lightweight neural networks, which are responsible for time feature extraction, vehicle density feature extraction, vehicle speed feature extraction, and the final fusion of these three features. This hierarchical feature extraction mechanism can fully explore the intrinsic connection between data of different dimensions and provide richer and more accurate input features for the prediction model. Among them, the time feature uses a matrix structure to store multi-dimensional time series features, the vehicle density feature takes into account the spatial occupancy differences of different vehicle models, and the vehicle speed feature integrates dynamic factors such as speed, acceleration and jerk. These feature extraction methods all reflect the in-depth analysis and modeling of the actual parking lot exit situation.

[0149] Secondly, the present invention uses a generative adversarial network (GAN) to expand vehicle traffic data. GAN can effectively simulate and generate synthetic data samples that are close to real data by establishing a generator network and a discriminator network. These expanded data not only increase the diversity of training samples, but also provide more sufficient data support for subsequent predictive modeling.

[0150] Thirdly, the present invention proposes a dynamic memory allocation optimization algorithm based on the memory demand characteristics of neural networks. The algorithm includes three parts: memory demand prediction equation, memory allocation equation and memory recovery equation. The memory demand prediction equation uses a linear weighted method to predict the memory demand of each sub-network based on the parameters, calculation amount, batch size and hidden layer size of the neural network. The memory allocation equation uses the principle of proportional allocation based on these prediction results and taking into account the task priority to reasonably allocate the total memory to each sub-network. This dynamic perception and adaptive allocation method can significantly improve the computing efficiency and resource utilization of the system.

[0151] Finally, based on the above prediction model, the present invention designs a comprehensive scheduling strategy that takes into account both exit channel adjustment and vehicle diversion guidance. The exit channel adjustment scheme dynamically monitors the current congestion situation and adjusts the number of open exit channels in a timely manner according to the prediction results to alleviate severe congestion. The vehicle diversion guidance scheme pushes the best exit selection suggestions to the drivers of vehicles about to enter the exit to guide the reasonable diversion of vehicles. This combination of active control and passive guidance can better balance the exit capacity and vehicle traffic efficiency, and ultimately achieve effective relief of the congestion problem at the parking lot exit.

[0152] A specific embodiment 1 of the present invention is provided below. The specific implementation of the steps in this embodiment 1 is described in detail as follows: The specific implementation of step S10 is to construct a tree-structured joint neural network. The joint neural network consists of four lightweight neural networks, which are responsible for time feature extraction, vehicle density feature extraction, vehicle speed feature extraction and feature fusion respectively.

[0153] The first lightweight neural network is used to extract time features, and its time feature extraction matrix can be expressed as:

[0154] ;

[0155] in, Indicates The time segment time characteristics, represents the number of time slices, Represents the dimension of time features. This matrix structure can effectively represent multi-dimensional time series features.

[0156] The second lightweight neural network is used to extract vehicle density features, and its vehicle density feature vector can be expressed as:

[0157] ;

[0158] in, , , Represent the number vectors of small cars, medium cars and large cars respectively, , , is the corresponding weight coefficient, is the density correction term. This weighted summation method can comprehensively consider the differences in space occupancy of different types of vehicles.

[0159] The third lightweight neural network is used to extract vehicle speed features, and its vehicle speed features can be expressed as:

[0160] ;

[0161] in, For the The instantaneous speed of the vehicle, is the acceleration, is the acceleration, To measure the time interval, , , is the feature weight coefficient. This method of integrating velocity, acceleration, and jerk can fully describe the vehicle's motion state and highlight the importance of recent data through exponential decay.

[0162] The fourth lightweight neural network is responsible for fusing the above three features, and its feature fusion function can be expressed as:

[0163] ;

[0164] in, is the activation function (using ReLU), , , is the feature weight, This method of weighted summation and nonlinear activation can effectively capture the correlation between different features.

[0165] The specific implementation of step S20 is to use a generative adversarial network (GAN) to expand the vehicle traffic data. The generation loss function of GAN can be expressed as:

[0166] ;

[0167] in, is the generator loss, Output result for the generator, is a random noise vector, For real traffic data, is the discriminator’s judgment result on the generated data, is the smoothing loss term, , , is the weight coefficient. This loss function not only ensures the authenticity of the generated data, but also ensures the smooth continuity of the time series data and improves the diversity of the generated data.

[0168] The specific implementation of step S30 is to use a memory allocation optimization algorithm to allocate physical memory and graphics processor memory to the lightweight neural network. The memory demand prediction equation can be expressed as:

[0169] ;

[0170] in, For the The memory requirement of a lightweight neural network, is the parameter quantity, is the amount of calculation, is the batch size, is the hidden layer size, , , , is the weight coefficient, is the prediction error term. This linear weighted model considering the key parameters of the neural network can accurately predict the memory requirements.

[0171] The memory allocation equation can be expressed as:

[0172] ;

[0173] in, For the The amount of memory allocated for a lightweight neural network, is the task priority coefficient, is the total number of lightweight neural networks, is the total available memory of the system. This proportional allocation method based on memory demand and task priority can ensure the fairness and efficient use of memory resources.

[0174] The specific implementation of step S70 is to establish a vehicle congestion prediction model based on the fusion features. The model adopts a long short-term memory network (LSTM) structure, and its prediction function can be expressed as:

[0175] ;

[0176] in, for The predicted value of congestion probability at time for The fusion characteristics of the moment, is the time step, is the prediction error term. The LSTM network can effectively capture the long-term and short-term dependencies of time series data, thereby making more accurate congestion predictions.

[0177] The specific implementation of step S80 is to generate a vehicle diversion scheduling plan based on the vehicle congestion prediction result, and its optimization objective function can be expressed as:

[0178] ;

[0179] in, To optimize the goal, For the The average travel time per exit, For the The queue length of the lane, , is the weight coefficient, , This multi-objective optimization method that takes into account both travel time and queue length can achieve a balance between alleviating congestion and improving travel efficiency.

[0180] In order to better understand and implement the present invention, the following provides Example 2 of a specific application scenario of the present invention: A large parking lot in the central area of ​​a city has an average daily vehicle flow of more than 5,000 vehicles, and severe congestion often occurs at the exit during peak hours on weekdays. In order to alleviate this problem, the parking lot decides to adopt the parking lot exit vehicle congestion prediction and scheduling method proposed in the present invention.

[0181] First, multiple surveillance cameras and geomagnetic detectors are arranged at the exit of the parking lot to collect vehicle traffic data. The surveillance camera can extract the location information, speed information and type information of the vehicle, while the geomagnetic detector can obtain the vehicle's travel time information. Combining this information forms the training data set required by the present invention.

[0182] In order to further enrich the training data, the parking lot has also established a data expansion system based on the Generative Adversarial Network (GAN). The system consists of a generator network and a discriminator network. The generator network takes a random noise vector as input and outputs simulated vehicle traffic data. The discriminator network compares the differences between these simulated data and real vehicle traffic data, and feeds the difference results back to the generator network to optimize the parameters of the generator network. Through this generative adversarial training method, the parking lot has greatly increased the amount and diversity of training data, laying a good foundation for subsequent neural network training.

[0183] Next, the parking lot constructed a tree-structured joint neural network. Figure 2 As shown, the joint neural network consists of four lightweight neural networks, which are responsible for time feature extraction, vehicle density feature extraction, vehicle speed feature extraction and feature fusion respectively.

[0184] The first is the temporal feature extraction network. This network uses a three-layer convolutional layer and a two-layer fully connected layer structure. The input is the vehicle traffic data within a 30-minute time window, and the output is the temporal feature matrix. . The specific form is as follows:

[0185] ;

[0186] in, Indicates The time segment time characteristics, represents the number of time slices, Dimensions representing temporal features. These features include the average number of vehicles, variance, peak value, and periodicity.

[0187] The second network is the vehicle density feature extraction network. This network also uses a three-layer convolution and two-layer fully connected structure. The input is the number and type of vehicles in a 30-minute time window, and the output is the vehicle density feature vector. . The calculation formula is:

[0188] ;

[0189] in, , , Represent the number vectors of small cars, medium cars and large cars respectively, , , is the corresponding weight coefficient, This weighted summation method can comprehensively consider the differences in space occupancy of different types of vehicles.

[0190] The third network is the vehicle speed feature extraction network. The structure of this network is the same as the previous two networks. The input is the vehicle speed data within a 30-minute time window, and the output is the vehicle speed feature. . The calculation formula is:

[0191] ;

[0192] in, For the The instantaneous speed of the vehicle, is the acceleration, is the acceleration, To measure the time interval, , , is the feature weight coefficient. This method of integrating velocity, acceleration, and jerk can fully describe the vehicle's motion state and highlight the importance of recent data through exponential decay.

[0193] The fourth network is responsible for fusing these three features. This network adopts a four-layer fully connected layer structure, with the input being the output of the first three networks and the output being the final fused feature vector. . The calculation formula is:

[0194] ;

[0195] in, is the activation function (using ReLU), , , is the feature weight, This method of weighted summation and nonlinear activation can effectively capture the correlation between different features.

[0196] In order to optimize the operating efficiency of this joint neural network, the parking lot also adopts a dynamic memory allocation algorithm. The algorithm consists of three parts: memory demand prediction equation, memory allocation equation and memory recovery equation. Figure 3 As shown, the memory allocation of four lightweight neural networks is shown.

[0197] The specific form of the memory demand prediction equation is as follows:

[0198] ;

[0199] in, For the The memory requirement of a lightweight neural network, is the parameter quantity, is the amount of calculation, is the batch size, is the hidden layer size, , , , is the weight coefficient, is the prediction error term. This linear weighted model considering the key parameters of the neural network can accurately predict the memory requirements.

[0200] The specific form of the memory allocation equation is as follows:

[0201] ;

[0202] in, For the The amount of memory allocated for a lightweight neural network, is the task priority coefficient, is the total number of lightweight neural networks, is the total available memory of the system. This proportional allocation method based on memory demand and task priority can ensure the fairness and efficient use of memory resources.

[0203] Combining the above four lightweight neural networks, the parking lot constructs a tree-structured joint neural network to predict the congestion of exit vehicles. The prediction model adopts the long short-term memory network (LSTM) structure, and its prediction function can be expressed as:

[0204] ;

[0205] in, for The predicted value of congestion probability at time for The fusion characteristics of the moment, is the time step, is the prediction error term. The LSTM network can effectively capture the long-term and short-term dependencies of time series data, thereby making more accurate congestion predictions.

[0206] Based on the above prediction results, the parking lot has also designed a comprehensive scheduling strategy that takes into account both exit channel adjustment and vehicle diversion guidance. The exit channel adjustment plan dynamically monitors the current congestion situation and adjusts the number of open exit channels in a timely manner according to the prediction results to alleviate severe congestion. The vehicle diversion guidance plan pushes the best exit selection suggestions to vehicle drivers about to enter the exit to guide the reasonable diversion of vehicles. This combination of active control and passive guidance can better balance exit capacity and vehicle traffic efficiency.

[0207] Here is how the system performs in practice.

[0208] On a weekday, from 5 to 6 p.m., it is the peak time for the parking lot exit. During this period, the surveillance camera and geomagnetic detector collected the following data:

[0209] Table 1 Vehicle traffic data during peak hours

[0210] time Total number of vehicles Small car Mid-size car Large car Average speed 5:00-5:05 280 190 60 30 20 km / h 5:05-5:10 320 220 70 30 18 km / h 5:10-5:15 360 250 80 30 15 km / h 5:15-5:20 410 280 90 40 12 km / h 5:20-5:25 450 310 100 40 10 km / h 5:25-5:30 500 340 120 40 8 km / h 5:30-5:35 550 370 140 40 6 km / h 5:35-5:40 580 390 150 40 5 km / h 5:40-5:45 600 410 150 40 4 km / h 5:45-5:50 620 420 160 40 3 km / h 5:50-5:55 640 430 170 40 2 km / h 5:55-6:00 650 440 170 40 1 km / h

[0211] As can be seen from Table 1, during the hour from 5 to 6 o'clock, the traffic volume gradually increased, the number of small and medium-sized cars increased significantly, and the speed gradually decreased. These data reflect the dynamic characteristics of the gradual increase in exit congestion during this period. Figure 4 As shown in the figure, this is a double Y-axis line chart, which shows the changing trend of traffic volume and average speed over time. The left Y-axis represents the traffic volume (vehicles / 5min), and the right Y-axis represents the average speed (km / h).

[0212] By inputting these data into the aforementioned joint neural network model, the following congestion prediction results can be obtained:

[0213] Table 2 Vehicle congestion prediction results

[0214] time Blockage probability 5:00-5:05 0.2 5:05-5:10 0.3 5:10-5:15 0.4 5:15-5:20 0.6 5:20-5:25 0.7 5:25-5:30 0.8 5:30-5:35 0.9 5:35-5:40 0.9 5:40-5:45 0.95 5:45-5:50 0.97 5:50-5:55 0.98 5:55-6:00 0.99

[0215] As can be seen from Table 2, as time goes by, the congestion probability of the exit continues to rise and reaches a peak at around 5:40. This result provides an important basis for subsequent scheduling decisions.

[0216] Based on the above prediction results, the parking lot has taken the following diversion and scheduling measures:

[0217] 1. Dynamically adjust the number of export channels:

[0218] Keep 3 exits open before 5:15;

[0219] Between 5:15 and 5:30, only two exits will be open;

[0220] After 5:30, only one exit channel remains open.

[0221] 2. Push diversion suggestions to drivers: During peak hours, push detour routes to drivers entering the exit to guide them to choose other exits for diversion.

[0222] like Figure 5 As shown in the figure, the prediction results of congestion probability for different time periods and different vehicle types are displayed. The darker the color, the higher the probability of congestion, and the specific probability value is marked. Through this scheduling strategy that combines active regulation and passive guidance, the parking lot successfully alleviated the exit congestion problem during the peak period from 5 to 6 o'clock. The specific results are as follows:

[0223] Table 3 Scheduling effect evaluation

[0224] time Average travel time (min) Maximum queue length (m) Number of export channels 5:00-5:15 7 120 3 5:15-5:30 9 150 2 5:30-6:00 12 180 1

[0225] As can be seen from Table 3, after the dispatching measures were taken, the average travel time at the exit increased, but the maximum queue length was significantly alleviated. This shows that although the exit waiting time of individual vehicles increased, the overall travel efficiency and exit capacity were improved.

[0226] In addition, the parking lot also continuously monitors the execution of the scheduling plan and collects feedback data in a timely manner to continuously optimize the training of the joint neural network model. Every 24 hours, the parking lot will automatically update the model parameters to ensure the timeliness and accuracy of the prediction and scheduling decisions.

[0227] Through continuous optimization and implementation, the congestion problem at the exit of the parking lot has been effectively alleviated. The average travel time during peak hours is basically controlled within 10 minutes, and the maximum queue length is also controlled within 200 meters, which greatly improves the traffic efficiency of the exit. Drivers also generally reported that entering and exiting the parking lot has become smoother and more convenient.

[0228] It should be noted that the variables used in the present invention are specifically introduced in detail as shown in the following Table 4:

[0229] Table 4 Variable explanation table

[0230]

[0231] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for predicting and dispatching vehicle congestion at a parking lot exit, characterized in that: The following steps are involved: S10, constructing a tree-structured joint neural network, the joint neural network comprising four lightweight neural networks, a first lightweight neural network for extracting time features, a second lightweight neural network for extracting vehicle density features, a third lightweight neural network for extracting vehicle speed features, and a fourth lightweight neural network for feature fusion; S20, collecting vehicle traffic data at the parking lot exit, and using a generative adversarial network to perform data expansion on the vehicle traffic data to obtain a training data set; S30, optimizing the allocation of physical memory and graphics processor memory for the lightweight neural network based on a memory allocation optimization algorithm; S40, dividing the training data set into a plurality of training samples according to the time series, each of the training samples comprising vehicle travel time, vehicle quantity, and vehicle type; S50, inputting the training sample into the first lightweight neural network in the joint neural network to obtain a first eigenvector; S60, inputting the first feature vector into other lightweight neural networks in the joint neural network in sequence to obtain multiple prediction features; S70, establishing a vehicle congestion prediction model based on the multiple prediction features to obtain a vehicle congestion prediction result; S80, generating a vehicle diversion scheduling plan according to the vehicle congestion prediction result; S90: Execute the vehicle diversion scheduling plan, and collect execution results for updating the training data set.

2. A parking lot exit vehicle congestion prediction and dispatching method according to claim 1, characterized in that: The outputs of the first lightweight neural network, the second lightweight neural network, and the third lightweight neural network are respectively connected to the input end of the fourth lightweight neural network.

3. A parking lot exit vehicle congestion prediction and dispatching method according to claim 2, characterized in that: The first lightweight neural network, the second lightweight neural network, and the third lightweight neural network all adopt a three-layer convolutional layer and a two-layer fully connected layer structure, and the fourth lightweight neural network adopts a four-layer fully connected layer structure.

4. A parking lot exit vehicle congestion prediction and dispatching method according to claim 3, characterized in that: The steps of using a generative adversarial network for data expansion include: establishing a generator network and a discriminator network, wherein the generator network inputs a random noise vector and outputs simulated vehicle traffic data, and the discriminator network compares the difference between the simulated vehicle traffic data and the real vehicle traffic data, and feeds the difference result back to the generator network for optimizing the generator network.

5. A parking lot exit vehicle congestion prediction and dispatching method according to claim 4, characterized in that: The memory allocation optimization algorithm includes a memory demand prediction equation, a memory allocation equation and a memory recovery equation. The memory demand prediction equation determines the memory demand based on the parameter quantity and calculation amount of each lightweight neural network. The memory allocation equation allocates physical memory and graphics processor memory based on the memory demand. The memory recovery equation releases memory based on the task completion status of the lightweight neural network.

6. A parking lot exit vehicle congestion prediction and dispatching method according to claim 5, characterized in that: The step of acquiring the training data set includes: collecting parking lot exit camera image data, extracting vehicle position information, vehicle speed information and vehicle type information in the image, collecting parking lot exit geomagnetic detector data to obtain vehicle travel time information, and combining the vehicle position information, vehicle speed information, vehicle type information and vehicle travel time information to form training data; the training data set is updated using a sliding time window method, retaining historical data for the last 90 days, and updating the training of the joint neural network every 24 hours.

7. A parking lot exit vehicle congestion prediction and dispatching method according to claim 6, characterized in that: The training process of the joint neural network adopts a staged training method. First, the first lightweight neural network, the second lightweight neural network and the third lightweight neural network are trained separately. After completing the single feature extraction training, the parameters of the three lightweight neural networks are fixed, and the fourth lightweight neural network is trained to complete feature fusion.

8. A parking lot exit vehicle congestion prediction and dispatching method according to claim 7, characterized in that: The vehicle congestion prediction model adopts a long short-term memory network structure, with the input being the multiple prediction features and the output being the probability of vehicle congestion within the next 30 minutes; the vehicle diversion scheduling plan includes an exit channel quantity adjustment plan and a vehicle diversion suggestion plan, the exit channel quantity adjustment plan is used to dynamically adjust the number of open exit channels, and the vehicle diversion suggestion plan is used to push recommended exit selection information to vehicle drivers.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program instructions, and when the program instructions are executed in a computer, they are used to execute the parking lot exit vehicle congestion prediction and scheduling method according to any one of claims 1 to 8.

10. A parking lot exit vehicle congestion prediction and dispatching system, characterized in that: A computer-readable storage medium comprising the computer-readable storage medium of claim 9.

Images