Rail transit distributed lighting method, device and equipment based on machine vision

Abstract

The application discloses a kind of machine vision-based rail transit distributed lighting method, device and equipment, it is related to intelligent lighting technical field.The method comprises: collecting the multi-source monitoring data of each area of rail transit station, and carrying out real-time passenger flow analysis based on machine vision and future passenger flow prediction based on space-time graph neural network, obtain real-time passenger flow and predicted passenger flow information;Real-time passenger flow and predicted passenger flow information are input to lighting agent, space-time weighted fusion and microscopic autonomous decision are carried out, and preliminary lighting control action is obtained;All preliminary lighting control actions are summarized, global optimization is carried out using swarm intelligence optimizer, macroscopic control parameters are dynamically generated, and preliminary lighting control actions are re-evaluated;Optimized control action is executed, rail transit distributed lighting is realized, and swarm intelligence optimizer and lighting agent are co-evolved.Solve the technical problems of response lag, extensive control and lack of foresight in the prior art.

Description

Technical Field

[0001] This application relates to the field of intelligent lighting technology, and more specifically, to a method, apparatus, and equipment for distributed lighting in rail transit based on machine vision. Background Technology

[0002] As hubs of urban public transportation, rail transit stations are characterized by high passenger volume, complex spatial structures, and long operating hours. Their lighting systems not only affect passenger safety and visual comfort but are also a major source of energy consumption for the entire station. Traditional lighting control methods mainly include manual control and simple automated control based on timing and zoning. These methods have the following drawbacks:

[0003] 1) Response lag: The inability to perceive changes in passenger flow in real time leads to energy waste when passenger flow is sparse, and may affect passenger experience and safety due to insufficient lighting during peak passenger flow.

[0004] 2) Coarse control: It is usually controlled on a large area basis, which cannot achieve fine-grained on-demand lighting and is difficult to deal with the tidal phenomenon of passenger flow in local areas of the station.

[0005] 3) Lack of foresight: Unable to predict future passenger flow trends, only able to respond passively after passenger flow changes, unable to conduct forward-looking lighting guidance and resource pre-allocation.

[0006] With the development of artificial intelligence and Internet of Things technologies, some smart lighting solutions have begun to try to use sensor data for adjustment, but most of them are limited to the optimization of single points or local areas, lacking the ability to coordinate and optimize the overall performance of the entire site, and failing to effectively integrate the prediction information of future passenger flow.

[0007] Therefore, there is an urgent need for a distributed lighting method that combines real-time sensing, accurate prediction, micro-level autonomous decision-making, and macro-level global optimization to achieve efficient, comfortable, and intelligent operation of rail transit station lighting systems. Summary of the Invention

[0008] This application aims to provide a machine vision-based distributed lighting method, apparatus, and equipment for rail transit, which addresses the technical problems of response lag, coarse control, and lack of predictability in the prior art.

[0009] The first aspect of this application provides a machine vision-based distributed lighting method for rail transit, comprising:

[0010] Distributed data acquisition equipment is used to collect multi-source monitoring data from various areas of rail transit stations, and real-time passenger flow analysis based on machine vision and future passenger flow prediction based on spatiotemporal graph neural network are performed to obtain real-time passenger flow and predicted passenger flow information.

[0011] Real-time and predicted passenger flow information is input into the lighting intelligent agents in each area of ​​the rail transit station, and spatiotemporal weighted fusion and micro-autonomous decision-making are performed to obtain the preliminary lighting control actions of each area.

[0012] All initial lighting control actions are aggregated, and a swarm intelligence optimizer is used for global optimization to dynamically generate macro-control parameters. Based on the lighting agent, the initial lighting control actions are re-evaluated to generate optimized control actions.

[0013] It executes optimized control actions for each area of ​​rail transit stations, realizes distributed lighting for rail transit, and continuously promotes the collaborative evolution of the swarm intelligence optimizer and the lighting intelligence agent.

[0014] In one possible design approach, distributed data acquisition devices are used to collect multi-source monitoring data from various areas of the rail transit station, and real-time passenger flow analysis based on machine vision and future passenger flow prediction based on spatiotemporal graph neural networks are performed to obtain real-time passenger flow and predicted passenger flow information, including:

[0015] Distributed data acquisition devices are used to collect multi-source monitoring data from various areas of rail transit stations and preprocess the data to obtain corresponding standard data. The standard data includes passenger image data, train operation data, and basic environmental data.

[0016] Passenger image data is input into a machine vision-based multi-object detection model for real-time passenger flow analysis, generating a real-time passenger flow density heat map of rail transit stations;

[0017] Standard data is converted into a spatiotemporal graph and input into a passenger flow prediction model based on a spatiotemporal graph neural network to predict future passenger flow, generating a future passenger flow prediction matrix for rail transit stations within a preset time window.

[0018] Based on the regional division of rail transit stations, the real-time passenger flow density heat map and future passenger flow prediction matrix are converted into real-time passenger flow and predicted passenger flow information for each region, and sent to the corresponding lighting smart agent.

[0019] In one possible design approach, distributed data acquisition devices are used to collect multi-source monitoring data from various areas of the rail transit station, and the data is preprocessed to obtain corresponding standard data, including:

[0020] The distributed cameras in the distributed data acquisition device are used to collect video streams from various areas of the rail transit station. The structured data server in the distributed data acquisition device is used to collect train structured data from the ATS gateway. The distributed sensors in the distributed data acquisition device are used to collect raw sensor data from various areas of the rail transit station. The multi-source monitoring data, including video streams, train structured data, and raw sensor data, is sent to the edge computing gateway.

[0021] At the edge computing gateway, image preprocessing is performed on the raw image data of each video frame to obtain passenger image data, and data preprocessing is performed on the train structured data and raw sensor data to obtain train operation data and basic environmental data.

[0022] By integrating passenger image data, train operation data, and basic environmental data, standard data for each area of ​​the rail transit station is obtained.

[0023] In one possible design approach, passenger image data is input into a machine vision-based multi-object detection model for real-time passenger flow analysis, generating a real-time passenger flow density heatmap of the rail transit station, including:

[0024] At the edge computing gateway, passenger image data is input into a machine vision-based multi-object detection model to perform multi-object detection and obtain several passenger targets. The multi-object detection model is constructed based on the YOLOv7-tiny algorithm.

[0025] The DeepSORT algorithm is used to track each passenger target, and the unique ID, pixel coordinates and real-time speed of each passenger target are obtained.

[0026] The station map is divided into grids. Based on the unique ID, pixel coordinates and real-time speed of each passenger target, the number of passengers in each grid is counted in real time to obtain the passenger flow density. A real-time passenger flow density heat map of the rail transit station is generated and sent to the cloud server along with standard data.

[0027] In one possible design approach, standard data is converted into a spatiotemporal graph and input into a passenger flow prediction model based on a spatiotemporal graph neural network to predict future passenger flow, generating a future passenger flow prediction matrix for rail transit stations within a preset future time window, including:

[0028] On a cloud server, real-time passenger flow density heatmaps, as well as standard data on train operation and environmental baselines, are converted into spatiotemporal maps. These spatiotemporal maps are then input into a passenger flow prediction model based on a spatiotemporal map neural network. The passenger flow prediction model is constructed based on the STGCN algorithm.

[0029] Using a passenger flow prediction model, the graph structure features of the spatiotemporal graph, constructed from node features, edge features, and global event features, are extracted. Based on these graph structure features, future passenger flow is predicted, generating a future passenger flow prediction matrix for rail transit stations within a preset time window.

[0030] In one possible design approach, real-time and predicted passenger flow information is input into the lighting intelligent agents in each area of ​​the rail transit station. Spatiotemporal weighted fusion and micro-level autonomous decision-making are then performed to obtain preliminary lighting control actions for each area, including:

[0031] Real-time and predicted passenger flow information is input into the lighting smart body of each area of ​​the rail transit station, and the corresponding local state vector is generated in the lighting smart body.

[0032] Spatiotemporal attention weights are generated for the local state vector, and the local state vector is dynamically weighted and fused based on the spatiotemporal attention weights to obtain the global context vector;

[0033] The local state vector and the global context vector are fused through a gating mechanism to obtain the enhanced state vector of the lighting agent, and the enhanced state vector is then input into the policy network of the lighting agent.

[0034] Based on the preset action space and prospective reward function, the policy network is used to generate preliminary lighting control actions for the corresponding area according to the enhanced state vector, and the generated real-time experience is stored in the experience playback pool.

[0035] In one possible design approach, all initial lighting control actions are aggregated, a swarm intelligence optimizer is used for global optimization, macro-control parameters are dynamically generated, and based on the lighting agent, the initial lighting control actions are re-evaluated to generate optimized control actions, including:

[0036] On the cloud server, the enhanced state vectors, preliminary lighting control actions, and passenger experience index of all lighting agents are periodically aggregated to the swarm intelligence optimizer. The current global performance index baseline is calculated, and the fitness function of the swarm intelligence optimizer is defined based on the global performance index baseline. The swarm intelligence optimizer is constructed based on the PSO-GA algorithm.

[0037] The reward weight vector of the forward reward function is used as the macro-control parameter vector, encoded as the position vector of the swarm intelligent optimizer, and the swarm intelligent optimizer is initialized.

[0038] Based on the fitness function and a pre-built digital twin platform, a swarm intelligence optimizer is used to perform global optimization. The obtained global optimal particle is then decoded to obtain the optimal macroscopic control parameters.

[0039] The optimal macro-control parameters are sent to the lighting intelligent agents in each area of ​​the rail transit station;

[0040] In the lighting agent, the corresponding forward-looking reward function is reconstructed based on the optimal macro-control parameters, thus obtaining the reconstructed forward-looking reward function;

[0041] A small-scale random perturbation is performed within the neighborhood of the initial lighting control action to generate several candidate lighting control actions, thus obtaining an optimized action space.

[0042] Based on the optimized action space and the reconstructed prospective reward function, the policy network is used to generate optimized control actions for the corresponding region according to the enhanced state vector, and the generated real-time experience is stored in the experience replay pool.

[0043] In one possible design approach, based on the fitness function and a pre-built digital twin platform, a swarm intelligence optimizer is used for global optimization. The resulting globally optimal particle is then decoded to obtain the optimal macroscopic control parameters, including:

[0044] The Logistic map is chosen as the chaos generator to generate a chaotic sequence, and the chaotic sequence is mapped to the solution space to obtain each initial particle in the initial particle swarm.

[0045] In each iteration update process, the candidate macro-control parameters corresponding to each particle in the initial or previous iteration particle swarm are injected into the prospective reward function of the simulated lighting agent of the pre-built digital twin platform, and the corresponding fitness value is calculated based on the simulated environment using the fitness function.

[0046] After completing the fitness calculation for all particles, update the historical best position of the particles and the global best particle of the particle swarm based on the fitness values.

[0047] The Levy flight mechanism is introduced, and with mutation probability, several particles in the initial or previous iteration particle swarm are cross-crossed and mutated to obtain several mutated and cross-crossed particles.

[0048] A convergence factor is introduced to update the positions of all initial, previous iteration particles, mutated or crossover particles to obtain the updated particle swarm for the current iteration. Based on the simulation environment, the corresponding fitness value is calculated using the fitness function.

[0049] After completing the fitness calculation for all particles, the global optimal particle of the particle swarm is updated based on the fitness value until the number of iterations reaches the iteration threshold. The position vector of the global optimal particle updated in the last iteration is then decoded to obtain the optimal macroscopic control parameters.

[0050] A second aspect of this application provides a machine vision-based distributed lighting device for rail transit, comprising:

[0051] The data acquisition and passenger flow prediction unit is used to collect multi-source monitoring data from various areas of the rail transit station using distributed data acquisition equipment, and to perform real-time passenger flow analysis based on machine vision and future passenger flow prediction based on spatiotemporal graph neural network to obtain real-time passenger flow and predicted passenger flow information.

[0052] The micro-autonomous decision-making unit is used to input real-time and predicted passenger flow information into the lighting intelligent agents in each area of ​​the rail transit station, perform spatiotemporal weighted fusion and micro-autonomous decision-making, and obtain the preliminary lighting control actions of each area.

[0053] The global action optimization unit is used to summarize all preliminary lighting control actions, use a swarm intelligence optimizer to perform global optimization, dynamically generate macro-control parameters, and re-evaluate the preliminary lighting control actions based on the lighting agent to generate optimized control actions.

[0054] The distributed lighting and co-evolution unit is used to execute optimized control actions for various areas of rail transit stations, realize distributed lighting for rail transit, and continuously co-evolve the swarm intelligence optimizer and the lighting agent.

[0055] A third aspect of this application provides an electronic device, comprising:

[0056] A processor and a memory; the memory is used to store code instructions; the processor is used to run the code instructions, causing the electronic device to perform the machine vision-based distributed lighting method for rail transit as described in the first aspect.

[0057] Beneficial effects:

[0058] This application provides a machine vision-based distributed lighting method, device, and equipment for rail transit. Combining machine vision and edge computing, it can acquire passenger flow density in each micro-area in real time and accurately, providing a high-resolution data foundation for lighting control. By using a spatiotemporal graph neural network to fuse multi-source information such as train operation, it can accurately predict future passenger flow trends, enabling the lighting system to respond in advance and achieve effective passenger flow guidance and resource pre-allocation. It innovatively constructs a two-layer closed-loop control architecture of "macro-global optimization (swarm intelligence optimizer) combined with micro-autonomous decision-making (multi-agent reinforcement learning model)". The lighting agents make local autonomous decisions under the guidance of macro strategies, which not only ensures the achievement of global goals but also gives each area a high degree of flexibility and adaptability. Through dynamically adjustable macro-control parameters (reward weights), it achieves a dynamic balance of multiple goals such as energy saving, comfort, uniformity, and smoothness. Operators can flexibly adjust the optimization focus according to the needs of different time periods, and it has a collaborative evolution mechanism. The lighting agents and swarm intelligence optimizer continuously learn and optimize during operation, which can adaptively respond to changes in station operation modes and achieve long-term performance improvement of the system. Attached Figure Description

[0059] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This is a flowchart of a machine vision-based distributed lighting method for rail transit proposed in one embodiment of this application.

[0061] Figure 2 This is a schematic diagram of the structure of a machine vision-based distributed lighting device for rail transit, as proposed in one embodiment of this application.

[0062] Figure 3 This is a schematic diagram of the structure of a machine vision-based distributed lighting device for rail transit, as proposed in one embodiment of this application. Detailed Implementation

[0063] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0064] like Figure 1As shown in the figure, this application provides a machine vision-based distributed lighting method for rail transit, including:

[0065] 101. Use distributed data acquisition equipment to collect multi-source monitoring data from various areas of rail transit stations, and perform real-time passenger flow analysis based on machine vision and future passenger flow prediction based on spatiotemporal graph neural networks to obtain real-time passenger flow and predicted passenger flow information.

[0066] 102. Input real-time and predicted passenger flow information into the lighting intelligent agents in each area of ​​the rail transit station, perform spatiotemporal weighted fusion and micro-autonomous decision-making, and obtain the preliminary lighting control actions of each area;

[0067] 103. Summarize all preliminary lighting control actions, use a swarm intelligence optimizer to perform global optimization, dynamically generate macro-control parameters, and re-evaluate the preliminary lighting control actions based on the lighting agent to generate optimized control actions.

[0068] 104. Execute optimized control actions for each area of ​​the rail transit station to realize distributed lighting for rail transit, and continuously coordinate the evolution of the swarm intelligence optimizer and the lighting intelligence agent.

[0069] In one possible design approach, distributed data acquisition devices are used to collect multi-source monitoring data from various areas of the rail transit station, and real-time passenger flow analysis based on machine vision and future passenger flow prediction based on spatiotemporal graph neural networks are performed to obtain real-time passenger flow and predicted passenger flow information, including:

[0070] 1011. Using distributed data acquisition equipment, collect multi-source monitoring data from various areas of the rail transit station, and preprocess the data to obtain corresponding standard data; the standard data includes passenger image data, train operation data, and basic environmental data.

[0071] 1012. Input passenger image data into a machine vision-based multi-object detection model for real-time passenger flow analysis and generate a real-time passenger flow density heat map of rail transit stations.

[0072] 1013. Convert the standard data into a spatiotemporal graph and input it into a passenger flow prediction model based on a spatiotemporal graph neural network to predict future passenger flow and generate a future passenger flow prediction matrix for rail transit stations within a preset time window.

[0073] 1014. Based on the regional division of rail transit stations, convert the real-time passenger flow density heat map and future passenger flow prediction matrix into real-time passenger flow and predicted passenger flow information for each region, and send them to the corresponding lighting intelligent agent.

[0074] In one possible design approach, distributed data acquisition devices are used to collect multi-source monitoring data from various areas of the rail transit station, and the data is preprocessed to obtain corresponding standard data, including:

[0075] 10111. Use distributed cameras in the distributed data acquisition device to collect video streams from various areas of the rail transit station. Use the structured data server in the distributed data acquisition device to collect structured train data from the Automatic Train Supervision (ATS) gateway. Use distributed sensors in the distributed data acquisition device to collect raw sensor data from various areas of the rail transit station. Send the multi-source monitoring data, including video streams, structured train data, and raw sensor data, to the edge computing gateway.

[0076] In this embodiment, the distributed cameras continuously collect H.264 / H.265 encoded video streams from key areas such as platforms, passageways, and escalators at a rate of 15-25 frames per second;

[0077] The structured data server obtains the following structured data from the ATS gateway at a frequency of 1 second via industrial protocols such as OPC UA or Modbus TCP: train ID, current route, direction of travel, distance to the target platform, real-time speed, estimated arrival time, platform number, and current door status (open / closed).

[0078] Distributed sensors, such as light sensors and temperature and humidity sensors, are used to collect real-time ambient illuminance, color temperature, and temperature data for each area, which serve as the basis for calibration and decision support.

[0079] 10112. At the edge computing gateway, image preprocessing is performed on the raw image data of each video frame to obtain passenger image data, and data preprocessing is performed on the train structured data and raw sensor data to obtain train operation data and environmental basic data.

[0080] In this embodiment, the video frames are denoised (e.g., using Gaussian filtering), distortion corrected, and brightness normalized to improve the robustness of subsequent algorithms.

[0081] Data cleaning, format conversion, and magnitude normalization are performed on train structured data and raw sensor data to improve data quality;

[0082] 10113. Integrate passenger image data, train operation data, and basic environmental data to obtain standard data for each area of ​​the rail transit station.

[0083] In one possible design approach, passenger image data is input into a machine vision-based multi-object detection model for real-time passenger flow analysis, generating a real-time passenger flow density heatmap of the rail transit station, including:

[0084] 10121. At the edge computing gateway, passenger image data is input into a machine vision-based multi-object detection model for multi-object detection to obtain several passenger targets; the multi-object detection model is constructed based on the YOLOv7-tiny algorithm.

[0085] In this embodiment, a lightweight multi-target detection model (YOLOv7-tiny) is used to detect passenger heads or whole bodies in video frames in real time on the edge computing gateway, thereby obtaining several passenger targets.

[0086] 10122. Use a simple online and real-time tracking algorithm (DeepSORT) to track each passenger target and obtain the unique ID, pixel coordinates and real-time speed of each passenger target;

[0087] 10123. Divide the station map into grids (e.g., 0.5m x 0.5m). Based on the unique ID, pixel coordinates, and real-time speed of each passenger target, count the number of passengers in each grid in real time to obtain the passenger flow density (people / square meter), generate a real-time passenger flow density heat map of the rail transit station, and send the real-time passenger flow density heat map and standard data to the cloud server.

[0088] In one possible design approach, standard data is converted into a spatiotemporal graph and input into a passenger flow prediction model based on a spatiotemporal graph neural network to predict future passenger flow, generating a future passenger flow prediction matrix for rail transit stations within a preset future time window, including:

[0089] 10131. On a cloud server, the real-time passenger flow density heat map, as well as the standard data of train operation and environmental basic data, are converted into a spatio-temporal map, and the spatio-temporal map is input into a passenger flow prediction model based on a spatio-temporal graph neural network; the passenger flow prediction model is constructed based on the Spatio-Temporal Graph Convolutional Network (STGCN) algorithm.

[0090] In this embodiment, each logical area of ​​the station (such as platform A, passage B, and escalator C) is defined as a node in the graph. The edges between nodes are defined according to physical connectivity (such as passage connection) and functional correlation (such as platform and transfer passage). The initial weight of the edges can be set according to the passenger flow transfer probability analyzed from historical data.

[0091] 10132. Using a passenger flow prediction model, extract the graph structure features of the spatiotemporal graph, which are constructed from node features, edge features, and global event features. Based on these graph structure features, predict future passenger flow and generate a future passenger flow prediction matrix for rail transit stations within a preset future time window (e.g., 15 minutes in the future, with one time point per minute). M [One region, 15 time steps];

[0092] In this embodiment, the node characteristic is: past. N minutes (e.g.) N Within 30), the environmental baseline data for the real-time passenger flow density time series and standard data of each node in the real-time passenger flow density heatmap;

[0093] Edge features: connectivity and dynamic weights between nodes;

[0094] Global event characteristics: Train arrival / departure event sequences from standard train operation data acquired from ATS, used as externally injected pulse signals.

[0095] In one possible design approach, real-time and predicted passenger flow information is input into the lighting intelligent agents in each area of ​​the rail transit station. Spatiotemporal weighted fusion and micro-level autonomous decision-making are then performed to obtain preliminary lighting control actions for each area, including:

[0096] 1021. Input real-time passenger flow and predicted passenger flow information into the lighting intelligent agent of each area of ​​the rail transit station, and generate the corresponding local state vector in the lighting intelligent agent;

[0097] In this embodiment, for example, the local state vector:

[0098] Passenger flow density in this area at present, in the past, and at several future points in time;

[0099] The brightness and color temperature of current, past, and future lighting fixtures in this area;

[0100] Basic ambient illuminance in this area;

[0101] The estimated arrival time and status of train events affecting this area;

[0102] 1022. Generate the spatiotemporal attention weights of the local state vector, and dynamically weight and fuse the local state vector based on the spatiotemporal attention weights to obtain the global context vector;

[0103] This embodiment includes the following details:

[0104] Query generation: The agent inputs its local state vector into a fully connected layer to generate a "query vector";

[0105] Key-value pair generation: All other agents also input their local state vectors into their respective fully connected layers to generate "key vectors" and "value vectors". All "key vectors" and "value vectors" are pooled into a global information pool.

[0106] Attention score calculation: The agent calculates the dot product of its query vector and all key vectors, and normalizes it through the Softmax function to obtain the attention weights for all other regions. This weight quantifies the importance of other regions to the decision of the current region at the current moment.

[0107] Dynamic weighted fusion: The agent uses the calculated attention weights to perform a weighted sum of all value vectors to obtain a "global context vector";

[0108] 1023. The local state vector and the global context vector are fused through a gating mechanism to obtain the enhanced state vector of the lighting agent, and the enhanced state vector is input into the policy network of the lighting agent (such as an Actor network, a multilayer perceptron (MLP)).

[0109] In this embodiment, the enhanced state vector not only includes "what happened here", but also "what happened in which places in the world (in the station) is most relevant to me";

[0110] The policy network outputs a continuous action vector, such as [brightness adjustment value, color temperature adjustment value], or [+15%, -200K] which means increasing the brightness by 15% and decreasing the color temperature by 200K.

[0111] 1024. Based on the preset action space and prospective reward function, and according to the enhanced state vector, a policy network is used to generate preliminary lighting control actions for the corresponding area, and the generated real-time experience is stored in the experience playback pool.

[0112] In this embodiment, the formula for the forward-looking reward function is:

[0113]

[0114] In the formula, For reward value; The reward for energy saving is (-current total power / rated maximum power), a negative value for energy saving items; For comfort rewards, the value is (current illuminance, current passenger density). The higher the match between illuminance and density, the larger the value. Calculate future light environment smoothness reward H Within a step, the variance of the rate of change of brightness in this region and its neighboring regions; the smaller the variance, the smoother the change. The higher; As a reward for pre-guidance benefits, if a surge in passenger flow is predicted for a certain route and the lighting for that route is optimized in advance, a positive reward will be given. , For instant rewards; , As a forward-looking reward; The reward weights are used to form the subsequent macro-control parameter vector. When generating the initial lighting control action, the reward weights optimized in the previous step are used.

[0115] After the decision is generated, the lighting intelligence records a complete tuple of experiences. And store it in the local experience replay pool. This is the augmented state vector before the action is performed. To optimize control actions, This refers to the reward value calculated using a forward-looking reward function based on the current macroeconomic control parameters after the action is performed. The new enhanced state vector after the action is performed. This is a time-indicating quantity; the process continues, allowing the experience replay pool to continuously accumulate data reflecting real-world interactions.

[0116] In one possible design approach, all initial lighting control actions are aggregated, a swarm intelligence optimizer is used for global optimization, macro-control parameters are dynamically generated, and based on the lighting agent, the initial lighting control actions are re-evaluated to generate optimized control actions, including:

[0117] 1031. On a cloud server, the enhanced state vectors, preliminary lighting control actions, and passenger experience indices of all lighting agents are periodically aggregated to the swarm intelligence optimizer. The current global performance index baseline is calculated, and the fitness function of the swarm intelligence optimizer is defined based on the global performance index baseline. The swarm intelligence optimizer is constructed based on the Particle Swarm Optimization-Genetic Algorithm (PSO-GA) algorithm.

[0118] In this embodiment, the global performance metrics include:

[0119] Total energy consumption of the station in the past 5 minutes;

[0120] The current uniformity of illumination across the entire network is calculated as follows: For all areas with passengers, calculate the standard deviation of their illuminance values. The smaller the standard deviation, the higher the uniformity.

[0121] The current average visual comfort index for passengers is a comprehensive score based on illuminance, color temperature, and dynamic change rate.

[0122] The formula for the fitness function is:

[0123]

[0124] In the formula, For particles X The corresponding fitness value; For particles X The corresponding total energy consumption of the station; For particles X Corresponding overall network illumination uniformity; For particles X The corresponding average visual comfort index for passengers; For fitness weights; the fitness function is achieved through... The weighted combination creates a "value balance", and operations managers can adjust the weights to clarify the primary objectives at different times;

[0125] 1032. The reward weight vector of the forward-looking reward function As a macro-control parameter vector, it is encoded as the position vector of the swarm intelligent optimizer, and the swarm intelligent optimizer is initialized, where, i For particle indication;

[0126] 1033. Based on the fitness function and a pre-built digital twin platform, a swarm intelligence optimizer is used to perform global optimization. The obtained global optimal particle is then decoded to obtain the optimal macroscopic control parameters.

[0127] 1034. Send the optimal macro-control parameters to the lighting intelligent agents in each area of ​​the rail transit station;

[0128] 1035. In the lighting agent, based on the optimal macro-control parameters, the corresponding forward-looking reward function is reconstructed to obtain the reconstructed forward-looking reward function;

[0129] 1037. Perform a small-scale random perturbation within the neighborhood of the initial lighting control action to generate several candidate lighting control actions, thereby obtaining an optimized action space.

[0130] 1038. Based on the optimized action space and the reconstructed prospective reward function, the policy network is used to generate optimized control actions for the corresponding region according to the enhanced state vector, and the generated real-time experience is stored in the experience replay pool.

[0131] In one possible design approach, based on the fitness function and a pre-built digital twin platform, a swarm intelligence optimizer is used for global optimization. The resulting globally optimal particle is then decoded to obtain the optimal macroscopic control parameters, including:

[0132] 10331. Select the Logistic mapping as the chaos generator to generate a chaotic sequence, and map the chaotic sequence to the solution space to obtain each initial particle in the initial particle swarm.

[0133] The formula is:

[0134]

[0135] In the formula, For the first i+ 1. i One chaotic variable; The stability coefficient is typically 4. This sequence is ergodic and random, ensuring that the initial particle swarm is uniformly distributed in the solution space, avoiding getting trapped in local optima, which is superior to traditional random initialization.

[0136]

[0137] In the formula, The first in the particle swarm i An initial particle; For the first i One chaotic variable; These are the upper and lower bounds of the search space;

[0138] 10332. In each iteration update process, the candidate macro-control parameters corresponding to each particle in the initial or previous iteration particle swarm are injected into the prospective reward function of the simulated lighting agent of the pre-built digital twin platform, and the corresponding fitness value is calculated based on the simulated environment using the fitness function.

[0139] In this embodiment, in order to reduce the computational pressure on the cloud server, a lightweight neural network (such as MLP or LSTM) is used as an alternative model for the fitness function. The input is the macro-control parameters and the output is the global performance index. By learning the deep relationship between the macro-control parameters and the global performance index, fast and accurate fitness calculation is achieved.

[0140] 10333. After completing the fitness calculation for all particles, update the historical best position of the particles and the global best particle of the particle swarm based on the fitness values.

[0141] 10334. Introduce the Levy flight mechanism, and with mutation probability, perform cross-mutation on several particles in the initial or previous iteration of the particle swarm to obtain several mutated / crossed particles.

[0142] In this embodiment, crossover: generates a random number between (0, 1). If the crossover probability <Cross probability threshold Then, a crossover operation is performed, randomly selecting another particle. ( j ≠ i ,in, j (The particle indicator) is used as a cross partner, and arithmetic crossover is employed, with the formula as follows:

[0143]

[0144] In the formula, For the first t In the particle swarm of the next iteration A cross-shaped particle; For the first t The iteration of the ... The initial or previous iteration particle; For the first t Cross partners in the next iteration; for Levy Distributed random numbers; b for Levy Step length, and b ∈[1,2]; t This represents the current iteration number; The cross parameter is between (0,1);

[0145] Mutation: Generates a random number between (0,1) If the mutation probability <probability threshold Then, a number of particles are randomly selected to perform a mutation operation, as shown in the formula:

[0146]

[0147] In the formula, For the first t In the particle swarm of the next iteration A mutated particle; For the first t The iterations are distinct and ≠ i Three random particles; For variation parameters;

[0148] 10335. Introduce a convergence factor to update the positions of all initial, previous iteration particles, mutated or crossover particles to obtain the updated particle swarm for the current iteration, and calculate the corresponding fitness value based on the simulation environment using the fitness function.

[0149] The formula is:

[0150]

[0151] In the formula, For the first t +1 ,t The iteration of the ... i The velocity of each particle; This is the collaboration coefficient, which can be set to the same or different values ​​to adjust the influence weight. A random number in the range [0,1]. Inertia weights constructed for the convergence factor For the first t The initial, mutated, or crossover particles of the next iteration; For the first t The iteration of the ... i The best historical position of each particle; For the first t The globally optimal particle in the next iteration;

[0152]

[0153] In the formula, This represents the maximum number of iterations. , To adjust the parameters; It is the hyperbolic tangent function; This represents the maximum value of the convergence factor. This represents the minimum convergence factor.

[0154] 10336. After completing the fitness calculation of all particles, update the global optimal particle of the particle swarm based on the fitness value until the number of iterations reaches the iteration threshold. Decode the position vector of the global optimal particle updated in the last iteration to obtain the optimal macroscopic control parameters.

[0155] In one possible design approach, optimized control actions are executed for each area of ​​the rail transit station to achieve distributed lighting for the rail transit system. Continuous collaborative evolution of the swarm intelligence optimizer and the lighting agent is also implemented, including:

[0156] 1041. The distributed lighting control device receives optimized control actions from the cloud server (e.g., brightness adjusted to 85%, color temperature adjusted to 4000K), and through its built-in drive circuit (such as DALI, Zigbee or PLC communication module), it accurately sends the control commands to each LED lamp in its jurisdiction. The lamps then adjust their output power and spectrum to execute the action, thereby achieving real-time and precise control of the physical light environment.

[0157] 1042. Autonomous Evolution of Lighting Agents:

[0158] The evolution of a lighting agent is essentially a parameter optimization process for its decision-making core—the policy network, achieved through online or offline reinforcement learning training, specifically including:

[0159] The lighting agent incorporates a lightweight training module, which is activated during periods of low system load (such as early morning each day). The training process is as follows:

[0160] a. Randomly sample a batch of experience tuples from the experience replay pool;

[0161] b. Using the experience gained from sampling, calculate the policy gradient based on a pre-defined reinforcement learning algorithm (such as Proximal Policy Optimization (PPO)).

[0162] c. Using the calculated gradient, update the weight parameters of the policy network through the backpropagation algorithm;

[0163] Through this continuous "trial and error-learning" cycle, the policy network of the lighting agent can gradually learn how to make better decisions based on the state under the guidance of the current macro-control parameters, thereby achieving the autonomous evolution of its micro-decision-making ability.

[0164] 1043. Policy Evolution of Swarm Intelligence Optimizers:

[0165] The evolution of swarm intelligence optimizers does not refer to changes in their algorithms (such as PSO-GA), but rather to the dynamic adjustment of their optimization objectives, namely the fitness weights in the fitness function, to adapt to higher-level operational needs. Specifically, this includes:

[0166] 1) The cloud server continuously records and archives the results of each global optimization, including: macro-control parameter vectors, corresponding global performance indicators (total energy consumption, illumination uniformity, average passenger visual comfort index), timestamps, special event markers (such as weekdays, weekends, holidays, large-scale events), etc.

[0167] 2) A meta-learning strategy module runs on the cloud server, which learns and makes decisions based on long-term performance data:

[0168] a. Pattern recognition: Using clustering or time series analysis algorithms, identify different station operation modes (such as morning peak mode, evening peak mode, off-peak mode, night mode, and holiday mode).

[0169] b. Weighted Learning: For each operating mode, the meta-learning module analyzes historical data to find the fitness weight combination that maximizes the long-term comprehensive benefits of the mode (such as the weighted integral of energy saving and comfort).

[0170] c. Strategy Deployment: The meta-learning module determines the station's operating mode based on the current time, calendar, and real-time events, and dynamically injects the corresponding optimized fitness weights as a new "value balance" configuration into the fitness function of the swarm intelligence optimizer.

[0171] 3) The system provides a management interface that allows operations managers to manually adjust fitness weights based on real-time or anticipated management goals. For example, when facing power shortages, the weight of energy consumption items can be significantly increased; when hosting large-scale events, the weights of passenger comfort and light uniformity can be increased. These adjustments by managers are also recorded by the meta-learning module as part of future learning.

[0172] 1044. Co-evolutionary mechanism:

[0173] The evolution of lighting agents and swarm intelligence optimizers is not independent, but rather forms a tightly coupled co-evolutionary closed loop:

[0174] a. Macro-level guidance for micro-level: The swarm intelligence optimizer sets a clear direction and "values" for the evolution of all lighting agents (policy network training) by evolving their fitness weights. The lighting agents learn how to better serve the macro-level goal;

[0175] b. Micro-level feedback to macro-level: As the policy networks of lighting agents become stronger, their simulated behavior in digital twin platforms will increasingly approximate their optimal performance in the real world. This allows swarm intelligence optimizers to obtain more accurate fitness values ​​when evaluating alternative macro-level control parameters, thus enabling them to find the true global optimum more efficiently;

[0176] c. Spiral Ascent: The optimization of macro strategies provides better targets for the learning of micro agents, and the improvement of the capabilities of micro agents, in turn, improves the accuracy and efficiency of macro strategy optimization; the two promote each other and jointly drive the entire lighting system to achieve continuous performance improvement and spiral evolution in multiple dimensions such as energy saving, comfort, and safety, and finally reach a highly adaptive intelligent state.

[0177] Beneficial effects:

[0178] This application provides a machine vision-based distributed lighting method, device, and equipment for rail transit. Combining machine vision and edge computing, it can acquire passenger flow density in each micro-area in real time and accurately, providing a high-resolution data foundation for lighting control. By using a spatiotemporal graph neural network to fuse multi-source information such as train operation, it can accurately predict future passenger flow trends, enabling the lighting system to respond in advance and achieve effective passenger flow guidance and resource pre-allocation. It innovatively constructs a two-layer closed-loop control architecture of "macro-global optimization (swarm intelligence optimizer) combined with micro-autonomous decision-making (multi-agent reinforcement learning model)". The lighting agents make local autonomous decisions under the guidance of macro strategies, which not only ensures the achievement of global goals but also gives each area a high degree of flexibility and adaptability. Through dynamically adjustable macro-control parameters (reward weights), it achieves a dynamic balance of multiple goals such as energy saving, comfort, uniformity, and smoothness. Operators can flexibly adjust the optimization focus according to the needs of different time periods, and it has a collaborative evolution mechanism. The lighting agents and swarm intelligence optimizer continuously learn and optimize during operation, which can adaptively respond to changes in station operation modes and achieve long-term performance improvement of the system.

[0179] like Figure 2 As shown, this application provides a machine vision-based distributed lighting device for rail transit, including:

[0180] The data acquisition and passenger flow prediction unit 201 is used to collect multi-source monitoring data from various areas of the rail transit station using distributed data acquisition equipment, and to perform real-time passenger flow analysis based on machine vision and future passenger flow prediction based on spatiotemporal graph neural network to obtain real-time passenger flow and predicted passenger flow information.

[0181] The micro-autonomous decision-making unit 202 is used to input real-time passenger flow and predicted passenger flow information into the lighting intelligent agents of each area of ​​the rail transit station, perform spatiotemporal weighted fusion and micro-autonomous decision-making, and obtain the preliminary lighting control actions of each area.

[0182] The global action optimization unit 203 is used to summarize all preliminary lighting control actions, use a swarm intelligence optimizer to perform global optimization, dynamically generate macro-control parameters, and re-evaluate the preliminary lighting control actions based on the lighting agent to generate optimized control actions.

[0183] The distributed lighting and co-evolution unit 204 is used to execute optimized control actions for each area of ​​the rail transit station, realize distributed lighting for rail transit, and continuously co-evolve the swarm intelligence optimizer and the lighting agent.

[0184] The distributed lighting device for rail transit based on machine vision provided in this application embodiment can perform the above-described method and technical solution. Its principle and beneficial effects are similar, and will not be described again here.

[0185] like Figure 3 As shown, this is an electronic device provided in an embodiment of this application. The electronic device includes:

[0186] The memory 301 and the processor 302 communicate with each other via a communication bus 303.

[0187] The memory 301 is used to store code instructions.

[0188] The processor 302 is used to run the code instructions, causing the electronic device to execute the machine vision-based distributed lighting method for rail transit provided in the embodiments of this application.

[0189] The aforementioned communication bus 303 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus 303 can be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, only one thick line is used in the figure, but this does not indicate that there is only one bus or one type of bus. The communication interface is used for communication between the aforementioned terminal and other devices. The memory 301 can include random access memory (RAM), or it can include non-volatile memory, such as at least one disk storage device. Optionally, the memory 301 can also be at least one storage device located remotely from the aforementioned processor 302.

[0190] The processor 302 mentioned above can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0191] This application also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps in the method disclosed in this application.

[0192] This application also provides a computer program product that, when run on an electronic device, causes a processor to execute the steps in the method disclosed in this application.

[0193] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0194] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, electronic devices, and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0195] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0196] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0197] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0198] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0199] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A machine vision-based distributed lighting method for rail transit, characterized in that, include: Distributed data acquisition equipment is used to collect multi-source monitoring data from various areas of rail transit stations, and real-time passenger flow analysis based on machine vision and future passenger flow prediction based on spatiotemporal graph neural network are performed to obtain real-time passenger flow and predicted passenger flow information. Real-time and predicted passenger flow information is input into the lighting intelligent agents in each area of ​​the rail transit station. Spatiotemporal weighted fusion and micro-level autonomous decision-making are then performed to obtain preliminary lighting control actions for each area, including: Real-time and predicted passenger flow information is input into the lighting smart body of each area of ​​the rail transit station, and the corresponding local state vector is generated in the lighting smart body. Spatiotemporal attention weights are generated for the local state vector, and the local state vector is dynamically weighted and fused based on the spatiotemporal attention weights to obtain the global context vector; The local state vector and the global context vector are fused through a gating mechanism to obtain the enhanced state vector of the lighting agent, and the enhanced state vector is then input into the policy network of the lighting agent. Based on the preset action space and prospective reward function, the policy network is used to generate preliminary lighting control actions for the corresponding area according to the enhanced state vector, and the generated real-time experience is stored in the experience playback pool. All initial lighting control actions are aggregated, and a swarm intelligence optimizer is used for global optimization to dynamically generate macro-control parameters. Based on the lighting agent, the initial lighting control actions are re-evaluated to generate optimized control actions, including: On the cloud server, the enhanced state vectors, preliminary lighting control actions, and passenger experience index of all lighting agents are periodically aggregated to the swarm intelligence optimizer. The current global performance index baseline is calculated, and the fitness function of the swarm intelligence optimizer is defined based on the global performance index baseline. The swarm intelligence optimizer is constructed based on the PSO-GA algorithm. The reward weight vector of the forward reward function is used as the macro-control parameter vector, encoded as the position vector of the swarm intelligent optimizer, and the swarm intelligent optimizer is initialized. Based on the fitness function and a pre-built digital twin platform, a swarm intelligence optimizer is used to perform global optimization. The obtained global optimal particle is then decoded to obtain the optimal macroscopic control parameters. The optimal macro-control parameters are sent to the lighting intelligent agents in each area of ​​the rail transit station; In the lighting agent, the corresponding forward-looking reward function is reconstructed based on the optimal macro-control parameters, thus obtaining the reconstructed forward-looking reward function; A small-scale random perturbation is performed within the neighborhood of the initial lighting control action to generate several candidate lighting control actions, thus obtaining an optimized action space. Based on the optimized action space and the reconstructed prospective reward function, the policy network is used to generate optimized control actions for the corresponding region according to the enhanced state vector, and the generated real-time experience is stored in the experience replay pool. It executes optimized control actions for each area of ​​rail transit stations, realizes distributed lighting for rail transit, and continuously promotes the collaborative evolution of the swarm intelligence optimizer and the lighting intelligence agent.

2. The machine vision-based distributed lighting method for rail transit according to claim 1, characterized in that, Distributed data acquisition devices are used to collect multi-source monitoring data from various areas of rail transit stations. Real-time passenger flow analysis based on machine vision and future passenger flow prediction based on spatiotemporal graph neural networks are then performed to obtain real-time and predicted passenger flow information, including: Distributed data acquisition devices are used to collect multi-source monitoring data from various areas of rail transit stations and preprocess the data to obtain corresponding standard data. The standard data includes passenger image data, train operation data, and basic environmental data. Passenger image data is input into a machine vision-based multi-object detection model for real-time passenger flow analysis, generating a real-time passenger flow density heat map of rail transit stations; Standard data is converted into a spatiotemporal graph and input into a passenger flow prediction model based on a spatiotemporal graph neural network to predict future passenger flow, generating a future passenger flow prediction matrix for rail transit stations within a preset time window. Based on the regional division of rail transit stations, the real-time passenger flow density heat map and future passenger flow prediction matrix are converted into real-time passenger flow and predicted passenger flow information for each region, and sent to the corresponding lighting smart agent.

3. The machine vision-based distributed lighting method for rail transit according to claim 2, characterized in that, Distributed data acquisition devices are used to collect multi-source monitoring data from various areas of rail transit stations, and the data is preprocessed to obtain corresponding standard data, including: The distributed cameras in the distributed data acquisition device are used to collect video streams from various areas of the rail transit station. The structured data server in the distributed data acquisition device is used to collect train structured data from the ATS gateway. The distributed sensors in the distributed data acquisition device are used to collect raw sensor data from various areas of the rail transit station. The multi-source monitoring data, including video streams, train structured data, and raw sensor data, is sent to the edge computing gateway. At the edge computing gateway, image preprocessing is performed on the raw image data of each video frame to obtain passenger image data, and data preprocessing is performed on the train structured data and raw sensor data to obtain train operation data and basic environmental data. By integrating passenger image data, train operation data, and basic environmental data, standard data for each area of ​​the rail transit station is obtained.

4. The machine vision-based distributed lighting method for rail transit according to claim 3, characterized in that, Passenger image data is input into a machine vision-based multi-object detection model for real-time passenger flow analysis, generating real-time passenger flow density heatmaps for rail transit stations, including: At the edge computing gateway, passenger image data is input into a machine vision-based multi-object detection model to perform multi-object detection and obtain several passenger targets. The multi-object detection model is constructed based on the YOLOv7-tiny algorithm. The DeepSORT algorithm is used to track each passenger target, and the unique ID, pixel coordinates and real-time speed of each passenger target are obtained. The station map is divided into grids. Based on the unique ID, pixel coordinates and real-time speed of each passenger target, the number of passengers in each grid is counted in real time to obtain the passenger flow density. A real-time passenger flow density heat map of the rail transit station is generated and sent to the cloud server along with standard data.

5. The machine vision-based distributed lighting method for rail transit according to claim 4, characterized in that, Standard data is converted into a spatiotemporal graph and input into a passenger flow prediction model based on a spatiotemporal graph neural network to predict future passenger flow, generating a future passenger flow prediction matrix for rail transit stations within a preset future time window, including: On a cloud server, real-time passenger flow density heatmaps, as well as standard data on train operation and environmental baselines, are converted into spatiotemporal maps. These spatiotemporal maps are then input into a passenger flow prediction model based on a spatiotemporal map neural network. The passenger flow prediction model is constructed based on the STGCN algorithm. Using a passenger flow prediction model, the graph structure features of the spatiotemporal graph, constructed from node features, edge features, and global event features, are extracted. Based on these graph structure features, future passenger flow is predicted, generating a future passenger flow prediction matrix for rail transit stations within a preset time window.

6. The machine vision-based distributed lighting method for rail transit according to claim 5, characterized in that, Based on the fitness function and a pre-built digital twin platform, a swarm intelligence optimizer is used for global optimization. The globally optimal particle is then decoded to obtain the optimal macroscopic control parameters, including: The Logistic map is chosen as the chaos generator to generate a chaotic sequence, and the chaotic sequence is mapped to the solution space to obtain each initial particle in the initial particle swarm. In each iteration update process, the candidate macro-control parameters corresponding to each particle in the initial or previous iteration particle swarm are injected into the prospective reward function of the simulated lighting agent of the pre-built digital twin platform, and the corresponding fitness value is calculated based on the simulated environment using the fitness function. After completing the fitness calculation for all particles, update the historical best position of the particles and the global best particle of the particle swarm based on the fitness values. The Levy flight mechanism is introduced, and with mutation probability, several particles in the initial or previous iteration particle swarm are cross-crossed and mutated to obtain several mutated and cross-crossed particles. By introducing a convergence factor, the positions of all particles are updated to obtain the updated particle swarm for the current iteration, and the corresponding fitness value is calculated using a fitness function based on the simulation environment. After completing the fitness calculation for all particles, the global optimal particle of the particle swarm is updated based on the fitness value until the number of iterations reaches the iteration threshold. The position vector of the global optimal particle updated in the last iteration is then decoded to obtain the optimal macroscopic control parameters.

7. A machine vision-based distributed lighting device for rail transit, applied to the machine vision-based distributed lighting method for rail transit as described in any one of claims 1-6, characterized in that, include: The data acquisition and passenger flow prediction unit is used to collect multi-source monitoring data from various areas of the rail transit station using distributed data acquisition equipment, and to perform real-time passenger flow analysis based on machine vision and future passenger flow prediction based on spatiotemporal graph neural network to obtain real-time passenger flow and predicted passenger flow information. The micro-autonomous decision-making unit is used to input real-time and predicted passenger flow information into the lighting intelligent agents in each area of ​​the rail transit station, perform spatiotemporal weighted fusion and micro-autonomous decision-making, and obtain the preliminary lighting control actions of each area. The global action optimization unit is used to summarize all preliminary lighting control actions, use a swarm intelligence optimizer to perform global optimization, dynamically generate macro-control parameters, and re-evaluate the preliminary lighting control actions based on the lighting agent to generate optimized control actions. The distributed lighting and co-evolution unit is used to execute optimized control actions for various areas of rail transit stations, realize distributed lighting for rail transit, and continuously co-evolve the swarm intelligence optimizer and the lighting agent.

8. An electronic device, characterized in that, include: Processor and memory; The memory is used to store code instructions; The processor is used to run the code instructions, causing the electronic device to perform the machine vision-based distributed lighting method for rail transit as described in any one of claims 1-6.

Images