A campus intelligent pick-up and drop-off management method and system

Abstract

The application relates to the field of campus wisdom systems, and discloses a campus wisdom pick-up and drop-off management method, which comprises the following steps: acquiring real-time observation data of a target vehicle in a parking lot; updating a belief state of an internal traffic state of the parking lot based on the real-time observation data and a preset system dynamic model, wherein the system dynamic model is coupled with a traffic flow model for simulating micro behaviors of vehicles; and based on the updated belief state, a partially observable Markov decision process optimization algorithm is used to decide a target berth and a preliminary guide path for the target vehicle. The application combines the partially observable Markov decision process with the cellular automaton micro traffic flow simulation, and combines a particle filter to update the belief state, so that highly intelligent prediction and decision of the parking lot dynamics are realized, and the problems of poor adaptability, easy local congestion or uneven resource allocation of the parking lot in a complex dynamic environment are effectively overcome.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of campus intelligent system, in particular to a campus intelligent pick-up and drop-off management method and system. BACKGROUND

[0002] The campus intelligent pick-up and drop-off management method and system belong to the technical field of intelligent traffic management. This technology aims to optimize the vehicle parking and traffic order in the campus, especially in the pick-up and drop-off area during the peak period of students' going to and coming from school, through information technology and automation means. The application goal is to improve the pick-up and drop-off efficiency, ensure traffic safety, and improve the campus environment. The technology in this field usually involves vehicle identification, parking space state monitoring, information guidance, and background management, and is widely used in various schools and educational parks with centralized pick-up and drop-off needs.

[0003] Traditional management methods often rely on fixed area division or manual on-site scheduling, and lack effective prediction and dynamic response ability for the dynamic traffic flow in the campus, especially the complex congestion situation formed by the concentrated arrival of vehicles during the peak period. This leads to that the parking space resource allocation is often based on immediate observation or simple rules, and it is difficult to achieve global optimization. Vehicles blindly search for parking spaces due to information asymmetry, which exacerbates local congestion and reduces traffic efficiency. At the same time, many existing systems do not handle the inherent uncertainty and potential errors in sensor data (such as misjudgment or omission of video identification of parking space occupation) finely, or rely too much on high-precision sensing devices for accuracy. This not only increases the deployment cost, but also makes it difficult to guarantee the robustness of decision-making and the reliability of the system under actual complex lighting or shielding conditions. SUMMARY

[0004] The purpose of the present application is to provide a campus intelligent pick-up and drop-off management method and system, which solves the problems of insufficient prediction and decision optimization ability in response to dynamic traffic flow in the campus, insufficient processing of uncertainty in perception information, and lack of long-term self-adaptive adjustment mechanism in the system model in the prior art.

[0005] To achieve the above purpose, the present application is implemented by the following technical scheme:

[0006] A campus intelligent pick-up and drop-off management method, comprising the following steps:

[0007] Obtaining real-time observation data of a target vehicle in a parking lot;

[0008] When a target vehicle, i.e. a vehicle requiring pick-up service, enters or is about to enter the parking lot, various sensing devices deployed in the parking lot are activated to collect diversified real-time observation data. These data not only include the vehicle identity and accurate arrival time obtained through the license plate recognition system at the entrance, but also cover the visual interpretation of the occupancy of each adjacent berth through the video monitoring system, as well as the preliminary quantification of the macro traffic parameters such as the traffic density and average speed of the main internal channels. This process aims to build a comprehensive environmental snapshot, the mechanism of which is to provide original input for subsequent state recognition and decision-making through multi-point and multi-dimensional data collection. The various sensing units, license plate recognition unit, video image analysis unit, work together to converge the collected information stream to the subsequent data processing and state estimation link.

[0009] updating the belief state of the internal traffic state of the parking lot based on the real-time observation data and a preset system dynamic model, the system dynamic model coupling a traffic flow model for simulating micro behaviors of vehicles;

[0010] updating the belief state of the overall traffic state of the parking lot based on the real-time observation data obtained in the previous step and the pre-constructed system dynamic model. The "belief state" here is not an absolute assertion of the parking lot condition, but a probabilistic description reflecting the system's comprehensive judgment of the real occupancy of each berth, the real congestion degree of each road section, and the accurate position and intention of the vehicles in the parking lot under the current information. The system dynamic model plays a key role in this link, which embeds a fine traffic flow model that is specifically used to simulate the micro driving behaviors of each vehicle in the parking lot, including acceleration, deceleration, following and reaction to other vehicles and road conditions. This coupling enables the system not only to understand the current observation, but also to predict how the traffic state will naturally evolve without new intervention. Its operation mechanism is rooted in probabilistic inference, which is usually realized through advanced estimation algorithms such as particle filtering: the algorithm uses the dynamic model to predict the prior distribution of the state, and then combines the current actual observation data to correct the distribution, thereby obtaining a posterior probability distribution closer to the true situation, i.e. the updated belief state. This belief state updating module is the core that connects the past and the future, which digests the uncertainty of observation data and integrates the inherent evolution logic of the traffic system, providing a solid and uncertainty-aware cognitive basis for subsequent intelligent decision-making.

[0011] based on the updated belief state, using a partially observable Markov decision process optimization algorithm to decide the target vehicle's target adjacent berth and preliminary guiding path;

[0012] After obtaining the accurate probabilistic cognition of the parking lot state, the system further updates the belief state based on this, and uses the optimization algorithm of the partially observable Markov decision process (POMDP) to decide the optimal target temporary parking space and a preliminary guiding path concept for the target vehicle currently in need of service. The POMDP framework is naturally suitable for sequential decision-making problems in the real world where the state cannot be accurately perceived. Here, the optimization algorithm, an online planner based on Monte Carlo tree search, will prospectively explore the long-term cumulative benefits brought by assigning different parking spaces and adopting different preliminary paths based on the current belief state. It will consider multiple evaluation dimensions such as expected waiting time, travel time, potential congestion, and parking space resource utilization balance, and assign corresponding weights. Its decision mechanism is to find a robust strategy that performs well under various real conditions through a large number of look-ahead simulations in an uncertain environment. The decision engine module receives the probabilistic global situation output by the belief state update module and combines it with the preset long-term optimization goal (reflected in the reward function) to output a parking space allocation scheme with guiding significance through complex search and evaluation.

[0013] According to the target temporary parking space and the real-time traffic conditions contained in the updated belief state, a detailed guiding path is planned;

[0014] Following the parking space decision, the system will plan a detailed and dynamically optimized guiding path for the target vehicle based on the target temporary parking space previously decided and again referring to the real-time traffic congestion conditions of each road segment in the parking lot contained in the belief state. This step is different from the preliminary path concept, as it will generate specific executable navigation instructions. The path planning algorithm, A* or its dynamic variants, will search on the topological graph representing the parking lot road network, but its core feature is that the passing cost of each road segment in the graph is not static but is updated in real time according to the belief state. If the belief state shows that a certain area currently has heavy traffic and slow speed, the passing cost of the relevant road segment leading to that area will increase accordingly. The mechanism is to achieve a dynamic optimal path selection considering real-time traffic conditions, ensuring that the vehicle is not only guided to a suitable parking space, but also follows a path that can effectively avoid congestion and shorten the time in the field under the current traffic conditions. The path planning module receives the parking space specified by the decision engine module and the real-time traffic information provided by the belief state update module, calculates and outputs a refined navigation scheme.

[0015] The target temporary parking space and the detailed guiding path are published to the user terminal or guiding device;

[0016] The system publishes the target parking stall number and detailed guidance path information calculated in the previous steps to the user terminal or guidance device in the parking lot through a suitable interface. The user terminal can refer to the driver's smartphone application, which receives personalized parking stall and path guidance; the guidance device includes variable information signs and LED display screens set at the entrance and key nodes of the parking lot, providing clear visual navigation prompts to the driver. The mechanism of this link is to convert the complex internal intelligent decision results into external user understandable and executable instructions, completing the closed loop from information processing to physical world intervention. The information publishing module as the output port of the system ensures that the optimization results can be effectively delivered to the end user, thereby guiding traffic behavior and achieving management goals.

[0017] Preferably, the step of updating the belief state of the internal traffic state of the parking lot based on the real-time observation data and the preset system dynamic model comprises:

[0018] The particle filter is used to update the belief state through particle prediction, particle weight update, and particle resampling steps.

[0019] In the particle prediction step, the state evolution of each particle is simulated by the traffic flow model, which is a cellular automaton model that updates the vehicle state on the cellular grid according to preset vehicle acceleration rules, safety distance maintenance rules, random slowing rules, and vehicle movement rules.

[0020] First is the particle prediction link. In this link, the system deduces the parking lot state represented by each particle at the previous time to the current time according to a preset system dynamic model. This deduction is not a simple extrapolation, but a dynamic simulation of the virtual parking lot world carried by the particle through a fine traffic flow model. The mechanism is to predict how each hypothetical state is most likely to evolve without new observation information.

[0021] In the particle prediction stage, the evolution of vehicle states within each particle is simulated by a cellular automaton (CA) model. The CA model discretizes the parking lot road network into a series of interconnected cell grids, and time is also advanced in discrete steps. Vehicles are treated as entities occupying one or more cells. The CA model updates the states of all vehicles on the grid according to a set of pre-defined microscopic behavior rules, which are designed to mimic the behavior decisions of real drivers. The specific rules include: acceleration rules for vehicles, which allow vehicles to gradually increase their speed up to the speed limit or desired speed of the road segment if conditions permit; safety distance maintenance rules, which force vehicles to maintain a sufficient safety distance from the vehicle in front or obstacles, and to decelerate if necessary to avoid collisions; random slowing rules, which introduce a certain probability for vehicles to slightly decelerate without obvious external reasons, to reflect the uncertainty and minor disturbances in driving behavior; and vehicle movement rules, which update the position of vehicles on the cell grid according to the final speed determined by the aforementioned rules.

[0022] Preferably, the step of deciding the target parking berth and preliminary guiding path for the target vehicle based on the updated belief state comprises:

[0023] The POMCP algorithm is used to construct a search tree starting from the root node representing the current belief state, through an iterative process of selection, expansion, simulation, and backpropagation.

[0024] In the simulation step, the CA model is used to simulate the evolution of the internal traffic state of the parking lot after the execution of different parking berth allocation and path guiding actions, and the cumulative return of the simulated path is evaluated according to the pre-set reward function.

[0025] In the process of deciding the target parking berth and preliminary guiding path for the target vehicle based on the updated belief state of the parking lot, an advanced POMDP optimization algorithm is used. Specifically, this optimization process is implemented through the POMCP algorithm. The core of the POMCP algorithm is that it starts from the belief state representing the current system's probabilistic cognition of the overall condition of the parking lot, and constructs and iteratively optimizes a search tree to prospectively explore the future impact of different decisions.

[0026] The POMCP algorithm starts with a root node representing the current belief state. Then, the search tree is gradually expanded and refined through a series of iterative "select-expand-simulate-backpropagate" loops. In the "select" phase, the algorithm chooses a most promising path to extend down the tree according to some exploration-exploitation policy. Upon reaching a leaf node or an under-explored node, the "expand" phase is entered, in which one or more new actions (assigning a specific parking spot to the target vehicle) are considered and added as new branches to the search tree.

[0027] In this phase, for a newly expanded action or a selected exploration path, one or more quick "look-ahead" simulations are performed to evaluate the long-term value of the action sequence. The mechanism is to collect statistical information about the pros and cons of a decision by simulating the possible future scenarios. A key component is that, in these simulation runs, the evolution of the traffic state within the parking lot, especially the changes in vehicle flow, queue length, and the formation or dissipation of potential congestion caused by different parking spot assignments and initial path guidance actions, are carefully modeled by a previously defined cellular automaton (CA) model. The CA model here serves as a dynamic "world model" or "prediction engine", which takes in a hypothetical action ("guide the vehicle to spot X") and the current simulation state, and then computes the next time step state of all vehicles within the parking lot according to its internal microscopic traffic rules (acceleration, avoidance, random slowing, movement).

[0028] After the CA model simulates a sequence of future states, a pre-defined reward function is used to evaluate the cumulative return or cost of the simulated path (i.e. a sequence of states and actions). The reward function encapsulates the management objectives, such as minimizing vehicle waiting time, reducing travel time, alleviating congestion level, or balancing parking spot utilization. The mechanism is to quantify various performance metrics observed in the simulation into a comprehensive evaluation value, so that the algorithm can compare the pros and cons of different decision sequences.

[0029] Preferably, the reward function is used to evaluate the immediate return of executing an action in the real state , which is calculated as:

[0030] ;

[0031] wherein , , , are pre-defined weight factors, is the vehicle the waiting time, for a vehicle under the action of the predicted travel time, for based on the real state and the action predicted system congestion index, for the parking space utilization imbalance degree.

[0032] Preferably, in the step of planning a detailed guidance path according to the target parking space under the decision and the real-time traffic condition contained in the updated belief state, the cost function of the edge of the path planning algorithm used The calculation formula is:

[0033] ;

[0034] Wherein, is the road segment connecting the parking lot internal node and node , is the current belief state, is the predicted travel time through the road segment based on the belief state predicted congestion level of the road segment based on the belief state , , is a preset weight coefficient.

[0035] Preferably, the step of obtaining the real-time observation data of the target vehicle in the parking lot further comprises quantifying the uncertainty of the observation data;

[0036] Wherein, the observation uncertainty of the parking space occupancy state is represented by an observation model , wherein is the real occupancy state of the parking space , is the observed occupancy state of the parking space , and the observation model includes the false alarm rate of the parking space being incorrectly identified as occupied and the false negative rate of the parking space being incorrectly identified as free .

[0037] Sensor readings are often affected by various factors such as environmental noise, self-precision limitations, and algorithm interpretation errors. In order to enable the subsequent decision-making process to consider this uncertainty and make more robust judgments, the system will give a probabilistic interpretation to the observed information. For example, the occupancy state of the target parking space ​Taking the occupancy status as an example, its observation uncertainty is described by a specialized observation model. This observation model essentially defines the actual occupancy status of the berth. The sensor system observes its state under certain conditions (actually idle or actually occupied). This represents the conditional probability of a specific condition (identified as idle or identified as occupied). More specifically, this model is typically characterized by two key parameters: one is the false alarm rate. The first refers to the probability that a sensor incorrectly identifies a berth as occupied when it is actually vacant; the second is the false negative rate. This refers to the probability that a parking space is actually occupied by a vehicle, but the sensor fails to detect it and incorrectly identifies it as vacant. These probability parameters... and This can be achieved through calibration tests in a controlled environment based on sensor factory specifications, or through statistical learning and dynamic adjustment during long-term system operation by comparing the sensor readings with other higher-precision reference information (manual verification or fusion judgment of different types of sensors). The mechanism lies in assigning a clear probability to each possible observation error. When the system subsequently fuses these observation data for belief state updates, it no longer treats them as absolute "facts," but rather as "evidence" with a certain level of confidence. The raw observation information acquired by the data sensing module undergoes an uncertainty quantification process (or is directly processed by the observation model unit within the belief state update module) before being used by the belief state update module (particle filter). This allows the uncertainty information to be integrated into the probabilistic reasoning framework, enabling the system to more intelligently weigh and interpret imperfect sensor inputs.

[0038] Preferably, the vehicle acceleration rule, safe distance maintenance rule, random slowing rule, and vehicle movement rule of the cellular automata model are specifically NaSch model rules, including the following steps:

[0039] Acceleration steps: ;

[0040] Steps for maintaining a safe distance: ;

[0041] Random slowing down step: with a preset probability If the vehicle's current speed ,but ;

[0042] Vehicle moving steps: ;

[0043] in, For vehicles current speed of the vehicle, preset acceleration value, preset maximum allowable speed, vehicle number of free cells between the vehicle and the vehicle or obstacle in front of it, vehicle position of the vehicle.

[0044] Preferably, the method further comprises a model adaptation step:

[0045] Periodically collecting parking lot operation history data, the operation history data including actual travel time of the vehicle, actual occupancy duration of the parking space, and actual traffic flow parameters of each road section of the parking lot;

[0046] Based on the deviation between the collected operation history data and the algorithm internal model prediction results, adjusting the parameters of the cellular automaton model, the reward function weight factor of the partially observable Markov decision process model, or the false positive rate and false negative rate of the observation model.

[0047] To ensure that the intelligent drop-off management method can continuously adapt to the evolving campus traffic characteristics and maintain its operational efficiency, the system also integrates a dynamic model adaptation and parameter optimization mechanism. This mechanism is not set once, but is triggered periodically or when the system performance indicators are found to be drifting unexpectedly. The core of it is to first collect the actual operation history data of the parking lot through a data recording and analysis unit. These data constitute a valuable real-world behavior profile, specifically covering the actual total time spent by vehicles from the entrance to the designated parking space, the actual occupancy duration of each adjacent parking space, and the actual traffic flow parameters of each key road section inside the parking lot at different times. The traffic density and average travel speed obtained through video analysis are continuously monitored.

[0048] Subsequently, a dedicated evaluation and calibration module will compare these accumulated real-world data with the predicted outputs or behavioral performances of the internal core models under the same time period and similar conditions. The simulated traffic flow of a specific road segment by the cellular automaton model will be compared with the actual observed traffic time; the weight settings of the constituent items (penalties for waiting time and congestion) in the reward function of the partially observable Markov decision process will also be evaluated by analyzing the actual results to determine whether they truly guide the system to achieve the expected macro-management goals. When such comparisons reveal systematic deviations, i.e., there are persistent and significant differences between the model's predictions and reality, the parameter adjustment process will be initiated. Specifically, if the cellular automaton model's simulation of traffic flow does not match the actual observed congestion patterns or vehicle behavior, its internal parameters, such as the random slowing probability of a specific road segment or the expected maximum speed of a vehicle, will be fine-tuned based on the deviation. The mechanism is to make the microscopic traffic simulation more closely resemble real driving behavior and road condition responses. Similarly, if the actual running results show that, although a certain goal (minimizing waiting time) is given a high weight in the reward function, this goal has not been effectively improved, or the system's decision-making is too biased towards a certain indicator at the expense of other important aspects, these weight factors will be reweighted and corrected based on long-term performance feedback. The mechanism is to better align the system's decision-making logic with the manager's real intentions and multi-objective balancing needs. In addition, the observation model responsible for interpreting sensor inputs will periodically calibrate its internal false alarm rate and false negative rate parameters by comparing historical sensor readings with higher confidence reference data (parking space status from geomagnetic sensors or manually checked records). The mechanism is to continuously improve the modeling accuracy of the system's uncertainty in raw sensory information, enabling it to more accurately infer the true state from noisy observations. This closed-loop adjustment process based on historical data feedback enables the entire management method to have the ability of self-learning and continuous optimization, gradually reducing the gap between the model and reality, thereby ensuring the long-term stability and improvement of decision-making quality.

[0049] Preferably, before the step of using a partially observable Markov decision process optimization algorithm to decide the target vehicle's target parking space and preliminary guidance path based on the updated belief state, the method further comprises:

[0050] using a M(t) / G / c / K queuing theory model to perform macro-situation awareness on the parking lot, estimating the average number of queuing vehicles at the parking lot entrance and the average waiting time of vehicles ;

[0051] and using the average number of queuing vehicles or the average waiting time of vehicles as one of the component features of the belief state, or as an input reference for the partially observable Markov decision process optimization algorithm when evaluating the reward function.

[0052] Before the core step of deciding a specific parking slot and preliminary guiding path for the target vehicle, the method also introduces a macro-situation awareness step, aiming to obtain a preliminary quantitative evaluation of the overall busy degree of the parking lot entrance area. This step specifically adopts the classic M(t) / G / c / K queuing theory model. Queuing theory model can analyze vehicle arrival, waiting for service and departure from a statistical point of view, where M(t) represents the time-varying vehicle arrival rate (different arrival rates during the morning peak and lunch break), G represents the service time (time required to complete parking slot allocation and preliminary guidance) following a general probability distribution, c represents the number of parallel processing or service channels of the system, and K represents the maximum queue length or the total number of vehicles of interest that the system can accommodate. By applying this model, the system can estimate the key macro-congestion indicators at the parking lot entrance in real time, including the predicted average number of vehicles in the queue and the average waiting time of vehicles .

[0053] Through mathematical modeling of the vehicle arrival and service process, even in the case of being unable to accurately track each vehicle, the overall trend and severity of the entrance congestion can be effectively statistically predicted. These macro-congestion evaluation information output by the queuing theory model are then seamlessly integrated into the subsequent refined decision-making process: they can be directly used as the belief state of the feature-rich system, so that the starting "world view" of the subsequent partially observable Markov decision process when planning already contains the awareness of the entrance congestion; or, these estimated average queue length or waiting time can also be used as an important reference input for the partially observable Markov decision process optimization algorithm when evaluating the reward function corresponding to different decision schemes (i.e. different parking slot allocation), a parking slot allocation scheme that leads to a significant increase in entrance queue will be negatively affected in the reward evaluation.

[0054] A system based on the above pick-up and drop-off management method, comprising:

[0055] A data perception module for obtaining real-time observation data of a target vehicle in a parking lot;

[0056] A belief state updating module connected to the data perception module, for updating the belief state of the traffic state inside the parking lot based on the real-time observation data and a preset system dynamic model, the system dynamic model coupled with a traffic flow model for simulating vehicle micro-behavior;

[0057] A decision engine module, connected with the belief state updating module, is configured to determine a target berthing position and a preliminary guiding path for the target vehicle based on the updated belief state by using a partially observable Markov decision process optimization algorithm;

[0058] A path planning module, connected with the decision engine module and the belief state updating module, is configured to plan a detailed guiding path according to the target berthing position and real-time traffic conditions contained in the updated belief state.

[0059] An information publishing module, connected with the decision engine module and the path planning module, is configured to publish the target berthing position and the detailed guiding path to a user terminal or a guiding device.

[0060] In summary, the present application has the following at least one beneficial technical effect:

[0061] 1. The present application combines partially observable Markov decision process with cellular automaton micro-traffic flow simulation, and combines particle filter for belief state updating, which realizes highly intelligent prediction and decision of parking lot dynamics. This forward-looking berthing allocation and path guiding strategy can significantly improve the berthing turnover efficiency and vehicle smoothness in the campus pick-up area. Compared with the existing technology which mainly relies on static rules or instant reaction to the current state for management, the present application effectively overcomes the problems of poor adaptability, easy to cause local congestion or uneven resource allocation in complex dynamic environment.

[0062] 2. The present application quantifies the uncertainty of the observation data in the perception layer, especially establishes a probability model for the false alarm and missed alarm in the berthing state recognition. This makes the system fully consider the inherent defects of the sensing information when making state estimation and subsequent decision, so as to make more stable and reliable judgment. Unlike the existing technology which often simplifies the sensor error or relies on high-precision but expensive sensing devices to ensure data quality, the present application improves the fault tolerance ability of the system in real complex environment by effectively managing uncertainty, even relying on conventional sensors.

[0063] 3. The present application introduces a model self-adaptation and parameter dynamic optimization mechanism, which can periodically calibrate and adjust the internal key model according to the deviation between actual running data and model prediction. This continuous learning and self-improving ability ensures that the management method can adapt to the long-term evolution of campus traffic mode and the change of characteristics in specific period, and maintains its efficiency. Compared with the existing technology which rarely adjusts the model parameters once set or relies on artificial experience for periodic maintenance, the present application solves the limitation that the performance of the model is degraded over time or with environmental changes due to the solidification of the model.

[0064] 4. The application innovatively introduces a macro-situation awareness link based on queuing theory before micro-decision, to evaluate the overall busy degree and potential queuing pressure of the parking lot entrance area. This macro-level prediction information is effectively integrated into the subsequent belief state construction and refined decision-making process, enabling parking space allocation and path planning to consider both individual vehicle needs and overall parking lot load. Compared with existing technologies that often focus on individual vehicle optimal path guidance, ignoring entrance bottleneck effects and overall system throughput management, the application better solves the problem of increased entrance congestion or low overall system operation efficiency caused by local optimization. BRIEF DESCRIPTION OF DRAWINGS

[0065] Figure 1 is a method flowchart of the application;

[0066] Figure 2 is a system architecture diagram of the application. DETAILED DESCRIPTION

[0067] The following will be described in detail in combination with the accompanying Figure 1 - the accompanying drawings Figure 2 , to further illustrate the application.

[0068] The application provides a campus intelligent pick-up and drop-off management method, as shown in the accompanying drawings, the campus intelligent pick-up and drop-off management method comprises the following steps: Figure 1

[0069] S1: Obtain real-time observation data

[0070] This step aims to collect information from various sensing devices deployed in the campus parking lot. These information constitutes the basis for subsequent intelligent decision-making. Parking lot entrance is usually equipped with license plate recognition (LPR) cameras, when the target vehicle, i.e. the vehicle requiring pick-up and drop-off service, enters, the LPR system can automatically identify its license plate number and record the arrival timestamp. These information is the key to identify the vehicle identity and trigger the management process.

[0071] At the same time, inside the parking lot, especially in the area near the parking space and the main channel, video surveillance cameras will be installed. These cameras not only provide a global view, but more importantly, through the video analysis algorithm, they can preliminarily judge the occupancy status of each parking space, and identify whether there is a vehicle parked on the parking space. Video analysis can also monitor the traffic flow on the main road section, estimate the number of vehicles on the road section, and the dynamic parameters of the average driving speed.

[0072] ​In some embodiments, to improve the accuracy of the parking space status detection, a geomagnetic sensor can be buried under each parking space. The geomagnetic sensor can determine whether the parking space is occupied by sensing the disturbance of the vehicle metal to the geomagnetic field, which can serve as an effective supplement to the video analysis results, especially in poor light or with obstructions.

[0073] Furthermore, optionally, the system can use ultra-wideband (UWB) high-precision positioning technology. By equipping some frequently visiting vehicles or management personnel with UWB tags and arranging UWB base stations in the parking lot, the target vehicle or personnel can be precisely positioned to the centimeter level. This provides high-quality data for obtaining accurate vehicle trajectories, assisting in determining parking space occupancy, and verifying subsequent microscopic traffic models.

[0074] All the data collected by these sensors, including license plate information, parking space occupancy image features, road segment vehicle count, vehicle speed estimation, geomagnetic signals, and UWB position coordinates, constitute the original real-time observation data set.

[0075] In view of the uncertainty of the observation data, the present application performs quantitative processing. Taking the recognition of the parking space occupancy state by the video monitoring as an example, the observation model is defined. In the model, represents the true occupancy state (free, occupied) of the parking space, and is the state observed by the video system. This observation model characterizes the uncertainty through two key parameters: the false alarm rate and the missed alarm rate . The false alarm rate indicates the probability that the parking space is actually free but is incorrectly identified as occupied. The missed alarm rate indicates the probability that the parking space is actually occupied but is incorrectly identified as free. These probability parameters can be obtained in advance by evaluating the performance of the video analysis algorithm on a standard test set, or they can be dynamically calibrated during the system operation by comparing with higher-precision sensors (geomagnetic or manual verification). These uncertainty parameters are the basis for correctly evaluating the weight of the observation data in subsequent belief state updates. After the preliminary formatting and time stamp alignment of the various observation data collected in this step, they will be used as input for the next step of belief state updating. Through the fusion of multiple heterogeneous sensors, the parking lot dynamics are captured as comprehensively as possible, and the imperfections of the observation data themselves are quantified, laying a foundation for subsequent probabilistic reasoning and robust decision-making.

[0076] After the preliminary formatting and time stamp alignment of the various observation data collected in this step, they will be used as input for the next step of belief state updating. Through the fusion of multiple heterogeneous sensors, the parking lot dynamics are captured as comprehensively as possible, and the imperfections of the observation data themselves are quantified, laying a foundation for subsequent probabilistic reasoning and robust decision-making.

[0077] S2: Update the belief state of the internal traffic state of the parking lot

[0078] ​Due to noise and incompleteness in sensor observations, the system cannot directly obtain the precise true state of the parking lot. Therefore, this step employs particle filter (PF) technology, fusing real-time observation data and the system dynamic model to probabilistically estimate the current true traffic state of the parking lot, i.e., update the system's belief state. Belief state It is about all possible real states The probability distribution.

[0079] Particle filters maintain a set of weighted particles. To approximate this state of belief. Each particle This represents a complete assumption about the current actual state of the parking lot, including all adjacent parking spaces. Actual occupancy status (idle, occupied, expected when it will be available), all key road sections Real-time traffic conditions (accurate vehicle list, vehicle speed, road density), and all vehicles in the parking lot. The actual status (precise location, speed, assigned target berth), and the entrance waiting queue. Information (queue length, waiting time for each vehicle). It is the weight of the assumed state. This is the total number of particles, usually set to hundreds to thousands to balance estimation accuracy and computational cost.

[0080] The particle filter update process mainly includes three sub-steps: particle prediction, particle weight update, and particle resampling.

[0081] 2.1 Particle Prediction (Propagation): This sub-step, based on the system dynamic model, predicts the particle from the previous time step. Each particle Propagate forward to the current moment To obtain the predicted particles The system dynamic model mainly consists of two parts:

[0082] External event evolution: Simulating the impact of external events not directly controlled by the system on the state. Based on historical statistics and real-time predictions, the time-varying average vehicle arrival rate is calculated. A Poisson process stochastic model is used to predict whether new external vehicles will arrive at the entrance queue. Simultaneously, the average service time of each berth is considered. Statistical distribution (Exponential distribution, Erlang distribution) to simulate events where parked vehicles complete pick-up and drop-off services and leave the berth. These external events will change the entry queue information and berth occupancy information in the particle state.

[0083] Internal traffic state evolution (based on cellular automaton CA model): For each particle The evolution of the position and speed of vehicles already present inside the parking lot (whether they are driving to a parking space or looking for an exit) is simulated by a pre-defined microscopic traffic flow model, the cellular automaton (CA) model. The parking lot road network is discretized into a one- or two-dimensional grid of cells. The state of each cell at any time can be either free or occupied by a vehicle at a certain speed. The present invention employs the classical NaSch model as the underlying rule of the CA, which can be adjusted according to the characteristics of the parking lot. The specific rules are as follows and are applied in parallel to all simulated vehicles within a particle at each discrete time step:

[0084] Acceleration: Vehicle tries to accelerate where is the current speed (in units of cells / time step) of the vehicle, is a pre-defined acceleration value (typically 1 cell / time step 2 if ), is the maximum allowed speed on this road segment.

[0085] Slow down due to other cars: Vehicle slows down to avoid collision. Let be the number of free cells between the vehicle and the vehicle in front of it or a fixed obstacle (red light, parked vehicle). Then .

[0086] Randomization: Simulate the uncertainty of driving behavior. With a pre-defined randomization probability , if the current speed of the vehicle , then its speed is reduced by 1: . is a key calibration parameter, reflecting the cautiousness of the driver or the small perturbations of the road conditions.

[0087] Movement: Vehicle moves forward by the corresponding number of cells according to its final determined speed : where is the cell position of the vehicle . By simulating all vehicles within a particle through the CA model, the new position and new speed of these vehicles at the current time are obtained, thus updating the particle state regarding the road segment traffic conditions and the in-transit vehicle states​ The part.

[0088] 2.2 Particle Weight Update:

[0089] In obtaining the predicted particle set Then, it is necessary to consider the current time. Actual observation data obtained from step S1 To evaluate the plausibility of each predicted particle and update its weights. For each predicted particle... Calculate its generation of current observation likelihood This likelihood is based on the observation model defined in S1. If predicting particles medium berth The state is "idle", but the actual observation If it is "occupied", then the likelihood contribution of this part is Conversely, if a particle is "occupied" but observed to be "idle," then the likelihood contribution is... If they match, then they are respectively and .

[0090] The predicted particle is obtained by multiplying the likelihoods (assuming independent conditions) of all observable components of the particle state (all berth states, vehicle counts or speed observations for certain road sections). The total likelihood. Then, the particle New weights Its old weight It is directly proportional to this total likelihood:

[0091] ;

[0092] Finally, the new weights of all particles are normalized, so that... .

[0093] 2.3 Particle Resampling: To avoid particle degradation (i.e., a few high-weight particles dominate the entire particle set, while most particles have too small a weight and lose representativeness), resampling is usually performed after the weights are updated. Based on the updated weights... Extract with replacement from the current particle set A new particle is generated. High-weight particles have a higher probability of being selected multiple times, while low-weight particles are eliminated. Common resampling methods include systematic resampling and multinomial resampling. After resampling, the weights of all selected particles are typically reset to [value missing]. .

[0094] After the above three sub-steps, the new particle set is obtained represents the current time updated belief state .

[0095] In some embodiments, to improve the accuracy of macro situational awareness, the average number of queued vehicles at the entrance of the parking lot estimated by the M(t) / G / c / K queuing model and the average waiting time of vehicles are incorporated as features into the state representation of each particle , or as the basis for adjusting the internal parameters of the CA model (vehicle generation rate).

[0096] Through the particle filter, the system can generate a probabilistic description of the current overall state of the parking lot under observation data full of uncertainty, combined with the fine simulation of the internal dynamics of the traffic system (CA model). This belief state is closer to the true situation than any single sensor reading, providing a robust basis for subsequent intelligent decision-making, effectively overcoming the challenges brought by sensor errors and environmental randomness.

[0097] S3: Decision on the target parking space and preliminary guidance path

[0098] After obtaining the updated belief state at the current time , the goal of this step is to select an optimal target parking space and a preliminary guidance path concept for the newly arrived target vehicle (or the in-parking vehicle that needs to be re-decided). This decision-making process uses the optimization algorithm of the partially observable Markov decision process (POMDP), which is specifically implemented as a partially observable Monte Carlo planning (POMCP). POMCP is an online planning algorithm that finds the approximately optimal action by constructing and searching a forward-looking decision tree based on the current belief state.

[0099] The core of POMCP is Monte Carlo tree search (MCTS). When a decision needs to be made for , the system starts from the root node representing the current belief state , and performs MCTS iterations (U is a preset parameter, several hundred to several thousand times, depending on the available decision-making time). Each MCTS iteration contains the following four standard sub-steps:

[0100] 3.1 Selection: Starting from the root node of the MCTS tree, child nodes are recursively selected and traversed downwards according to a tree strategy called UCB1 (Upper Confidence Bound 1 applied to Trees). The UCB1 strategy balances the "exploitation" of known good actions with the "exploration" of less explored actions. The selection process continues until a leaf node is reached. This leaf node represents a history of a series of (simulated) actions that have been performed.

[0101] 3.2 Expansion: If the selected leaf node is reached... It has not been fully expanded (i.e., there are some legal actions that start from this historical state but have not yet been tried in the MCTS tree). If so, choose one of the actions that has not been tried. Expand upon this. (An action) Here it refers to the target vehicle. Assign a specific adjacent berth And the initial path to the berth. After expansion, it will be displayed in the tree. Add a new child node to represent the history-action pair. .

[0102] 3.3 Simulation / Rollout:

[0103] Starting from the newly expanded node (or, if the selected leaf node cannot be expanded further, from that leaf node), perform a fast simulation until the preset simulation depth is reached or a termination state is reached (the vehicle successfully parks and leaves, or the simulation time expires). First, a specific state needs to be sampled from the belief state associated with the current MCTS node (or its approximation, a specific state from the particle set). Start. Then, apply the action currently being considered. (or the action selected by a quick default strategy during the rollout phase) ) to this sampling state Next, after calling the cellular automaton (CA) model again to simulate the execution of the action, all vehicles inside the parking lot (including the target vehicle) will... The microscopic traffic flow evolution (including other vehicles in the venue) is used to obtain the next simulated state. This CA simulation process is the POMDP state transition function. The core manifestation of this is that during the simulation process, it will be based on a preset reward function. This is used to evaluate the immediate reward at each step or in the final state. In this invention, the reward function aims to minimize a total cost, and therefore its form is typically a negative value of the cost. A typical reward function... The definition is as follows:

[0104] ;

[0105] in:

[0106] It is a (approximate) real state during the simulation process. It refers to the action performed.

[0107] It is a vehicle The waiting time (the time spent waiting for allocation in the entrance queue, or the time spent waiting for the target berth to become available). It is the collection of all waiting vehicles.

[0108] It is a vehicle In action Guided by this information, the estimated travel time from the current location to the target berth is calculated. This time can be estimated using the number of steps taken in the CA simulation, or by combining the road length and simulated speed. It is the collection of all vehicles currently moving within the venue.

[0109] Based on simulated state and actions The predicted system congestion index can be quantified by statistically analyzing the number and total length of high-density (above the threshold) or low-speed (below the threshold) road segments, or the length of internal queues formed, after CA simulation.

[0110] This is a penalty term for uneven berth utilization, used to encourage the system to use all available berths more evenly, avoiding some berths being overloaded while others are idle. It can calculate the variance or Gini coefficient of the number of times each berth is allocated or the projected occupancy rate.

[0111] These are preset non-negative weighting factors used to balance the importance of different optimization objectives (reducing waiting time, shortening travel time, alleviating congestion, and balanced utilization). These weights can be adjusted according to the preferences of campus administrators. The entire simulation (rollout) process accumulates these immediate rewards (considering discount factors). ), to obtain the total discount return for this complete simulation. .

[0112] 3.4 Backpropagation:

[0113] The total return obtained from the simulation Starting from the expanded node (or the leaf node that performed the rollout), propagate backwards up the MCTS tree to the root node. All visited (historical, action) nodes along the path update their statistics: incrementing the visit count by one. And update its average return estimate (Q value). . .

[0114] After completing the preset After this iteration, the MCTS tree has fully explored and evaluated the potential value of different decision sequences. At this point, the system selects the root node (representing the current true decision point and belief). The action with the highest Q-value (or, in some comprehensive evaluation, the most frequently visited) in the next level child node. As the current target vehicle The optimal decision made. This action. Specifically, this means allocating an optimal target berth. A preliminary concept of a guided path (Points to the area where the berth is located or a key intersection).

[0115] The POMCP algorithm enables efficient online planning in uncertain environments (through belief state modeling). It assesses the long-term impact of different decisions through forward-looking Monte Carlo simulations (where the CA model plays a crucial "world model" role, predicting micro-traffic evolution), thus avoiding short-sighted decisions based solely on the current situation. This allows the system to make smarter and more predictive parking allocations, effectively balancing multiple objectives such as waiting time, travel time, congestion levels, and resource utilization, ultimately improving overall pick-up and drop-off efficiency and user experience.

[0116] S4: Plan detailed guidance routes

[0117] In step S3, for the target vehicle The optimal target berth was determined. Next, the task of this step is to determine the target parking space and the real-time traffic conditions within the parking lot (based on the updated belief state). (Reflection) plans a route for a vehicle from its current location (usually the parking lot entrance, or its current location within the lot) to the target parking space. Detailed and dynamic driving routes .

[0118] The path planning employs a classical graph search algorithm, A* (A-star) algorithm or its variant D*Lite algorithm (especially suitable for dynamic environment). The road network of the parking lot is abstracted as a graph, where the key locations of intersections and parking space entrance points are the nodes of the graph, and the lanes or road segments connecting these nodes are the edges of the graph.

[0119] The core lies in the dynamic setting of the cost function of the edges For a road segment connecting node and node , its traversal cost is not just the static distance or free-flow time, but dynamically considers the current traffic congestion situation. The cost function can be specifically defined as:

[0120]

[0121] Where:

[0122] is the current belief state obtained from the S2 step.

[0123] is the predicted travel time through the road segment based on the current belief state . This time can be estimated by analyzing the vehicle density and average speed on the road segment represented by the particles in the belief state, which are obtained by CA simulation. If the belief indicates that the road segment is congested, the will increase accordingly.

[0124] is the congestion level or the additional traversal cost resulting therefrom of the road segment based on the current belief state . This can also be extracted from the statistical information (density, queue length) about the traffic flow of the road segment in the belief state and mapped to a quantitative congestion cost.

[0125] , are preset non-negative weight coefficients for balancing the two goals of minimizing travel time and avoiding congestion. These coefficients can be adjusted according to actual needs.

[0126] The path planning algorithm (A*) will use this dynamic cost function to search for the path with the minimum total cost from the current location of the vehicle to the target parking space in the parking lot road network graph. Since the cost function reflects the traffic situation (perceived through the belief state) in real time, the planned path It will intelligently avoid the road segments that are currently congested or are predicted to be congested in the near future, and choose a route that is less congested.

[0127] The planned detailed path It is usually represented as a sequence of ordered nodes or segments. Then, the system can convert this path into navigation instructions that are easy for the driver to understand, such as "turn left into B zone lane in 100 meters ahead" or "drive along the current lane to P035 parking spot".

[0128] It ensures that the vehicle is not only provided with a target parking spot, but also an optimized path to reach that spot, taking into account real-time dynamic traffic conditions. By dynamically adjusting the path cost, the vehicle can effectively avoid congestion, reduce unnecessary waiting and detours, and further shorten the total stay time in the parking lot, improving traffic efficiency. This is a good complement to the macro parking spot decision of S3.

[0129] S5: Publish guidance information and perform management

[0130] Once the detailed guidance path and the target parking spot are determined, this step is responsible for effectively delivering this information to the driver and recording relevant decisions for subsequent analysis and system state update.

[0131] Information publishing can be done through various channels:

[0132] Entrance information screen: At the entrance of the parking lot, a large screen can display the target parking spot number and preliminary area guidance assigned to the vehicle that has just entered.

[0133] In-parking guidance screen: Electronic guidance screens are set up at key intersections or diversion points within the parking lot, which can display dynamic turning arrows or path instructions based on the vehicle's current location (if available) and target parking spot.

[0134] Parental application (APP): If the parent uses the matching mobile APP, the system can push the assigned parking spot number, detailed navigation instructions (even in the form of a map) directly to their mobile phone.

[0135] Vehicle navigation system (if integrated): For vehicles equipped with integrated vehicle navigation devices, navigation information can be sent directly to the car machine.

[0136] In addition to information publishing, this step also includes:

[0137] System state recording: record the decisions made by the system for the vehicle this time (assigned parking , planned path ) and the relevant state information (beliefs , observations ) at the time of decision making are recorded into the system log. These data are crucial for the subsequent system performance evaluation, problem tracing, and model optimization.

[0138] Updating the action to be executed: the current decision action (i.e. a vehicle is assigned to a specific stall and starts driving along the planned path) is considered as an executed “macro action”, which information is taken as input to affect the belief state updating process (the particle prediction part in S2 step) of the next decision cycle (when another vehicle arrives).

[0139] This constitutes the bridge between intelligent decision making and actual traffic management, ensuring that the optimized results of the algorithm are accurately and timely delivered to the user, thus guiding his behavior. Meanwhile, the comprehensive data recording provides the basis for the continuous monitoring, learning, and improvement of the system.

[0140] Model adaptation and parameter optimization:

[0141] In order to enable the method described in the present invention to maintain high efficiency in the long term and adapt to potential changes in the campus traffic pattern, the system also includes a mechanism for model adaptation and parameter optimization. This step is usually performed offline periodically (daily, weekly) or when a decrease in system performance is detected.

[0142] The main contents include:

[0143] Running data collection and analysis: the system continuously collects and stores long-term running history data. These data include but are not limited to: the actual waiting time of vehicles from the entrance to the assigned stall, the actual driving time of vehicles in the field, the actual occupancy time of the stall, the actual traffic flow parameters (density and speed obtained through video analysis) of each road segment at different times, and the compliance of drivers to the guidance path (if available).

[0144] Performance evaluation and bias analysis: compare the actual running data collected with the prediction results of the internal models of the algorithm at the corresponding time. Compare the road segment travel time simulated by the CA model with the actual observed travel time; compare the predicted values (expected waiting time) of each component in the POMDP reward function with the actual waiting time. Identify the links with large prediction bias or leading to suboptimal decisions.

[0145] Model parameter calibration:

[0146] CA model parameters: adjust the key parameters of the CA model according to the differences between the actual traffic flow data and the CA simulation output, such as the random slowing probability of different road segments or different time periods , the maximum expected speed of vehicles Automated calibration techniques can be used to minimize the gap between simulation and reality.

[0147] Observation model parameters: compare the parking stall occupancy state identified by the contrast sensor (video) with the higher confidence reference data (magnetic sensor readings or manual spot check results), periodically re-estimate and update the false positive rate in the observation model and the false negative rate .

[0148] POMDP reward function weights: according to the long-term running effect and the change of management target (more emphasis on reducing extreme waiting time, or more emphasis on improving parking stall turnover rate), the weight factor of each item in the reward function can be adjusted . Learn the optimal weight from historical data based on reinforcement learning.

[0149] Policy network retraining: the decision-making strategy of POMDP is learned by deep reinforcement learning method (DeepQ-Network, DRQN), so the accumulated running data (state / belief, action, actual reward, sequence of next state / belief) can be used as new training samples for offline retraining or fine-tuning of policy network and / or value network, in order to obtain better decision-making ability.

[0150] The campus intelligent pick-up and drop-off management system described below can be correspondingly referred to the campus intelligent pick-up and drop-off management method described above.

[0151] Please refer to the attached Figure 2 , the present application also provides a kind of campus intelligent pick-up and drop-off management system, comprising:

[0152] Data perception module, for obtaining the real-time observation data of target vehicle in parking lot;

[0153] Belief state updating module, connected with the data perception module, for updating the belief state of the internal traffic state of the parking lot based on the real-time observation data and the preset system dynamic model, the system dynamic model is coupled with the traffic flow model for simulating the micro behavior of vehicle;

[0154] Decision engine module, connected with the belief state updating module, for updating the belief state based on the updated belief state, using a partially observable Markov decision process optimization algorithm, for the target vehicle to decide target temporary parking stall and preliminary guide path;

[0155] Path planning module, connected with the decision engine module and the belief state updating module, for planning detailed guide path according to the target temporary parking stall of the decision and the real-time traffic conditions contained in the updated belief state; ​

[0156] An information publishing module, connected with the decision engine module and the path planning module, is configured to publish the target berthing position and the detailed guiding path to a user terminal or a guiding device

[0157] The system of the embodiment can be used to execute the method embodiments described above, and has similar principles and technical effects, which will not be described here again.

[0158] Although the embodiments of the present application have been shown and described, it is to be understood that the application is not limited to these embodiments. It will be obvious to those skilled in the art that various changes, modifications, substitutions and alterations can be made thereto without departing from the spirit and scope of the present application, which should be limited only by the appended claims and their equivalents.

Claims

1. A smart campus shuttle management method, characterized in that, Includes the following steps: Obtain real-time observation data of target vehicles in the parking lot; The belief state regarding the traffic state inside the parking lot is updated based on the real-time observation data and the preset system dynamic model, which is coupled with a traffic flow model used to simulate the micro-behavior of vehicles. Based on the updated belief state, a partially observable Markov decision process optimization algorithm is used to determine the target vehicle's target berth and initial guidance path. Based on the target berth in the decision and the real-time traffic conditions included in the updated belief state, a detailed guidance route is planned; The target berth and the detailed guidance route are published to the user terminal or guidance device; The steps for updating the belief state of the traffic conditions inside the parking lot include: The belief state is updated using a particle filter through particle prediction, particle weight update, and particle resampling steps. In the particle prediction step, the state evolution of each particle is simulated by the traffic flow model, which is a cellular automaton model. The vehicle state on the cellular grid is updated according to preset vehicle acceleration rules, safe distance maintenance rules, random slowing rules, and vehicle movement rules. The steps for determining the target vehicle's nearest parking space and initial guidance path include: The partially observable Monte Carlo programming algorithm is used to construct a search tree starting from the root node representing the current belief state through an iterative process of selection, expansion, simulation, and backpropagation. In the simulation step, a cellular automata model is used to simulate the evolution of traffic conditions inside the parking lot after different parking space allocations and path guidance actions are executed, and the cumulative reward of the simulated path is evaluated according to a preset reward function.

2. The campus smart pick-up and drop-off management method according to claim 1, characterized in that, In the step of planning a detailed guidance route based on the target berth determined by the decision and the real-time traffic conditions included in the updated belief state, the cost function of the edge in the path planning algorithm employed is... The calculation formula is: ； in, To connect internal nodes of the parking lot and nodes The section of road, For the current state of belief, For belief-based states Predicted route The estimated travel time, For belief-based states Predicted road sections The level of congestion, , These are preset weighting coefficients.

3. The campus smart pick-up and drop-off management method according to claim 1, characterized in that, After the step of acquiring real-time observation data of target vehicles in the parking lot, the method further includes quantifying the uncertainty of the observation data; Among them, berths The observation uncertainty of the occupancy status is addressed through the observation model. It means that among them berth The actual occupancy status, berth The observation model includes the false alarm rate at which berths are incorrectly identified as occupied, and the occupancy status is monitored. The false alarm rate of berths being incorrectly identified as idle. .

4. The campus smart pick-up and drop-off management method according to claim 1, characterized in that, The method also includes a model adaptation step: Historical data on parking lot operation is collected periodically. This historical data includes the actual driving time of vehicles, the actual occupancy time of parking spaces, and the actual traffic flow parameters of each road segment in the parking lot. Based on the deviation between the collected historical data and the prediction results of the algorithm's internal model, the parameters of the cellular automata model, the weighting factor of the reward function of the partially observable Markov decision process model, or the false alarm rate and false negative rate of the observation model are adjusted.

5. The campus smart pick-up and drop-off management method according to claim 1, characterized in that, Before the step of using a partially observable Markov decision process optimization algorithm to determine the target vehicle's target berth and initial guidance path based on the updated belief state, the method further includes: The M(t) / G / c / K queuing theory model is used to perform macro-situational awareness of the parking lot and estimate the average number of vehicles queuing at the parking lot entrance. and average waiting time for vehicles ; and the average number of vehicles in the queue Or the average waiting time It can be used as one of the constituent features of the belief state, or as an input reference for the partially observable Markov decision process optimization algorithm when evaluating the reward function.

6. A system based on the pick-up and drop-off management method of claim 1, characterized in that, include: The data sensing module is used to acquire real-time observation data of target vehicles in the parking lot; The belief state update module, connected to the data perception module, is used to update the belief state of the traffic state inside the parking lot based on the real-time observation data and the preset system dynamic model. The system dynamic model is coupled with a traffic flow model for simulating the micro-behavior of vehicles. The decision engine module, connected to the belief state update module, is used to determine the target berth and preliminary guidance path for the target vehicle based on the updated belief state, using a partially observable Markov decision process optimization algorithm. The route planning module, connected to the decision engine module and the belief state update module, is used to plan a detailed guidance route based on the target berth in the decision and the real-time traffic conditions contained in the updated belief state. The information publishing module, connected to the decision engine module and the path planning module, is used to publish the target berth and the detailed guidance path to the user terminal or guidance device.

Images