Traffic signal management and control method and device based on AI edge computing

Abstract

The application discloses a traffic signal control method and device based on AI edge calculation, relates to the technical field of communication, and comprises the following steps: real-time interaction is carried out between a roadside edge node and a vehicle-mounted terminal, and a traffic right at an intersection is split into a space-time resource slot; when multiple vehicles request the same slot, an occupation conflict of the same space-time resource slot is detected, a blockchain auction is started, and a fair distribution is realized; adjacent edge nodes share the result and realize collaborative optimization through distributed machine learning, a signal control instruction signed by a number is generated, and signal light switching control is realized. The technical problems that fixed timing in traffic signal control does not match actual traffic changes, trunk green wave coordination is difficult to adapt to dynamic traffic changes, network delay and fluctuation of cloud control, and vehicle priority passage easily causes traffic signal disorder are solved, and the technical effects that the traffic efficiency of priority vehicles and ordinary vehicles mixed driving is improved, the real-time reliability of traffic signal control is improved, the collaborative control ability of adjacent intersections is enhanced, and the pressure of cloud control is reduced are achieved.

Description

Technical Field

[0001] This invention relates to the technical field of traffic signal control, specifically to a traffic signal control method and apparatus based on AI edge computing. Background Technology

[0002] Traditional traffic signal control systems often employ timed control, inductive control, or regional adaptive control based on centralized computing platforms, such as SCATS and SCOOT systems. These systems struggle to adapt to dynamically changing traffic demands, exhibiting significant limitations, particularly in handling sudden high-priority vehicle passage. Currently, to ensure the rapid passage of emergency vehicles such as fire trucks and ambulances, priority passage schemes based on radio frequency identification (RFID) or remote priority requests are commonly used. This typically involves directly intervening in intersection signals to forcibly switch to a green light for a specific direction. However, such methods have significant drawbacks. First, prioritizing emergency vehicles often comes at the expense of traffic flow efficiency in other phases of the intersection and even adjacent intersections, easily leading to local or even regional signal timing misalignment and traffic flow disorder. Second, due to the lack of global coordination and resource reservation mechanisms, multiple emergency vehicles requesting priority passage simultaneously or within similar timeframes may experience resource competition and path conflicts, delaying critical tasks. Finally, traditional centralized priority control modes suffer from high response delays and single-point-of-failure risks, failing to meet the demands of modern urban traffic management requiring high real-time performance and reliability.

[0003] Therefore, current technologies have several technical problems, including mismatch between fixed timing and actual traffic flow changes, difficulty in adapting green wave coordination on main roads to dynamic traffic flow changes, latency and fluctuations in cloud control networks, and the potential for traffic signal disruptions caused by vehicle priority passage. Summary of the Invention

[0004] This application provides a traffic signal control method and device based on AI edge computing, which solves the technical problems in existing traffic signal control such as mismatch between fixed timing and actual traffic flow changes, difficulty in adapting green wave coordination on trunk roads to dynamic traffic flow changes, latency and fluctuations in cloud control networks, and traffic signal disorder caused by priority vehicle passage. It achieves the technical effects of improving the passage efficiency and real-time reliability of traffic signal control when priority vehicles and ordinary vehicles are mixed, enhancing the collaborative control capability of adjacent intersections, and reducing the pressure on cloud control.

[0005] This application provides a traffic signal control method based on AI edge computing. The method includes: deploying edge nodes at traffic intersections, wherein the edge nodes integrate communication units, TPU accelerators, and multimodal sensors, enabling real-time acquisition of traffic flow data and execution of local decisions; vehicle-mounted terminals collecting the vehicle's priority status and driving path, generating encrypted passage demand messages, wherein the passage demand messages at least include vehicle identification, the expected sequence of target intersections to be passed, and the estimated time window for arrival at each intersection; a detector or RSU device forwarding the passage demand messages to the edge nodes at the target intersections, wherein after receiving the passage demand messages, the edge nodes divide the intersection right-of-way into continuous spatiotemporal resource slots according to the time dimension, wherein each slot defines the lane and time period that can be occupied; comparing the current traffic flow data with all other traffic flow data; and other methods. The system receives traffic request messages from vehicles, detects conflicts in the same spatiotemporal resource slots, and outputs a list of conflicting slots and a set of associated vehicles. Based on this list, for the conflicting spatiotemporal resource slots, edge nodes initiate a lightweight blockchain auction to generate auction results, including the winning vehicle and its precisely allocated spatiotemporal slot. Adjacent edge nodes share the auction results and coordinate their analysis using distributed machine learning to generate signal control commands, including phase timing commands, phase duration commands, and phase switching commands. Edge nodes use a random number generator to generate a dynamic key, sign the signal control commands, and distribute the signed commands to associated vehicles and intersection traffic light controllers for traffic light switching control.

[0006] In a possible implementation, after receiving the traffic demand message, the edge node splits the intersection right-of-way into continuous spatiotemporal resource slots according to the time dimension. Each slot defines the lane and time period that can be occupied. This includes: extracting an estimated time window from the traffic demand message to determine the time allocation range of the intersection right-of-way; dividing the time allocation range of the intersection right-of-way into time units according to a preset time granularity; combining each time unit with the lane space range to generate a spatiotemporal resource slot with a unique identifier; and classifying the spatiotemporal resource slots according to the priority marker in the vehicle identification to generate a set of spatiotemporal resource slots.

[0007] In a possible implementation, the passage request messages of all vehicles are compared to detect conflicts in the occupancy of the same spatiotemporal resource slots, and a list of conflicting slots and a set of associated vehicles are output. This includes: parsing the spatiotemporal resource slot requested by each passage request message; establishing a spatiotemporal resource slot occupancy status table to record the requesting vehicles for each spatiotemporal resource slot; traversing the spatiotemporal resource slot occupancy status table to identify the spatiotemporal resource slots requested by multiple passage request messages; and recording the spatiotemporal resource slots requested by multiple passage request messages and their requesting vehicle information as a list of conflicting slots and a set of associated vehicles.

[0008] In a possible implementation, based on the conflict slot list and associated vehicle set, for conflicting spatiotemporal resource slots, the edge node initiates a lightweight blockchain auction to generate auction results, including: setting auction parameters according to the conflict level of each spatiotemporal resource slot in the conflict slot list; sending bidding requests containing conflicting spatiotemporal resource slot information to vehicles in the associated vehicle set; receiving bidding responses submitted by each vehicle according to its priority status; processing the bidding responses according to preset bidding rules to determine the winning vehicle for each conflicting spatiotemporal resource slot; and recording the allocation relationship between the winning vehicle and the corresponding spatiotemporal resource slot as the auction result.

[0009] In a possible implementation, the bidding response is processed according to a preset bidding rule to determine the winning vehicle for each conflicting spatiotemporal resource slot. This includes: conducting a first round of open bidding for the conflicting spatiotemporal resource slots using a progressively decreasing price approach, and updating the current best bid in real time; when multiple identical best bids are detected, a second round of sealed bidding is initiated; based on the results of the second round of sealed bidding, the winning vehicle is determined according to the second-highest price principle; and the binding relationship between the winning vehicle and the corresponding spatiotemporal resource slot is written into the distributed ledger.

[0010] In a possible implementation, adjacent edge nodes share the auction results and perform coordinated analysis through distributed machine learning to generate signal control instructions. This includes: encapsulating the auction results, which contain the binding relationship between the winning vehicle and the spatiotemporal resource slot, into a coordination data packet; sending the coordination data packet to adjacent edge nodes through a secure channel; after verifying the coordination data packet, the adjacent edge nodes extract the auction result information from it; and based on the auction result information and local traffic data, generating a coordination scheme through a distributed machine learning model and encoding the coordination scheme into signal control instructions.

[0011] In a possible implementation, a coordination scheme is generated through distributed machine learning based on the auction result information and local traffic data. This includes: each edge node training a local model based on local traffic flow data and the received auction results; sharing model parameters through a secure aggregation protocol while protecting the privacy of the original data, aggregating the model parameters of multiple edge nodes, and generating a regional coordination model; and using the regional coordination model to evaluate different signal control schemes and select the optimal coordination scheme.

[0012] In a possible implementation, the regional coordination model is used to evaluate different signal control schemes and select the optimal coordination scheme, including: establishing an evaluation function for each edge node that includes traffic efficiency indicators, energy consumption cost indicators, and fairness indicators; calculating the evaluation function scores of different signal control schemes based on the vehicle allocation information in the auction results; and obtaining the optimal coordination scheme by iteratively calculating to make the evaluation function of each edge node reach a Nash equilibrium state.

[0013] In a possible implementation, the evaluation function of each edge node is brought to a Nash equilibrium state through iterative calculation, including: each edge node proposes initial signal control parameters based on the auction results and local traffic data; the initial control parameters and evaluation function calculation results of adjacent nodes are exchanged, and the impact of unilateral adjustment of control parameters on the overall evaluation function is calculated; when any edge node cannot improve its own evaluation function by unilaterally adjusting control parameters, it is determined that a Nash equilibrium state has been reached.

[0014] This application also provides a traffic signal control device based on AI edge computing. The device includes: an edge node deployment module for deploying edge nodes at traffic intersections, wherein the edge nodes integrate communication units, TPU accelerators, and multimodal sensors, enabling real-time acquisition of traffic flow data and execution of local decisions; a passage message generation module for onboard terminals to collect vehicle priority status and travel path, generating encrypted passage request messages, wherein the passage request message includes at least a vehicle identification identifier, the expected sequence of target intersections, and an estimated time window for arrival at each intersection; a passage message splitting module for detectors or RSU devices to forward the passage request message to the edge nodes at the target intersections, wherein after receiving the passage request message, the edge nodes split the intersection right-of-way into continuous spatiotemporal resource slots according to the time dimension, wherein each slot defines the allowed lane and time period for occupancy; and an occupancy conflict detection module. The system is used to compare the current traffic demand messages of all vehicles, detect conflicts in the occupancy of the same spatiotemporal resource slots, and output a list of conflicting slots and a set of associated vehicles. An auction result generation module is used to initiate a lightweight blockchain auction for conflicting spatiotemporal resource slots based on the list of conflicting slots and the set of associated vehicles, generating auction results including the winning vehicle and its precisely allocated spatiotemporal slot. A coordination analysis module is used for adjacent edge nodes to share the auction results, perform coordination analysis through distributed machine learning, and generate signal control instructions, including phase timing instructions, phase duration instructions, and phase switching instructions. A signal management module is used for edge nodes to generate dynamic keys using a random number generator, sign the signal control instructions, and distribute the signed signal control instructions to associated vehicles and intersection traffic light controllers for traffic light switching management.

[0015] This application proposes a traffic signal control method and device based on AI edge computing. By deploying AI edge nodes to collect traffic data in real time and execute local decisions, and combining this with encrypted passage request messages sent by vehicle terminals, the right-of-way at intersections is divided into spatiotemporal resource slots. After detecting conflicts in the occupancy of these slots, the edge nodes initiate a lightweight blockchain auction to resolve the conflict, determining the winning vehicle and its precise assigned slot. Adjacent nodes share the auction results through distributed machine learning, generating signal control commands, which are then dynamically signed and distributed to vehicles and traffic light controllers, achieving intelligent signal control. This addresses the technical problems in existing traffic signal control, such as the mismatch between fixed timing and actual traffic flow changes, the difficulty in adapting green wave coordination on arterial roads to dynamic traffic flow changes, latency and fluctuations in cloud control networks, and traffic signal disorder caused by priority vehicle passage. It achieves the technical effects of improving traffic efficiency and real-time reliability of traffic signal control when priority vehicles and ordinary vehicles share the road, enhancing the collaborative control capabilities of adjacent intersections, and reducing the pressure on cloud control. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings of the embodiments of this disclosure will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the apparatus according to the embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 This is a schematic diagram of the traffic signal control method based on AI edge computing provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the structure of a traffic signal control device based on AI edge computing, provided in an embodiment of this application.

[0019] Figure labeling: Edge node deployment module 10, passage message generation module 20, passage message splitting module 30, occupancy conflict detection module 40, auction result generation module 50, coordination analysis module 60, signal control module 70. Detailed Implementation

[0020] To further illustrate the technical means and effects adopted by the present invention in order to achieve the intended purpose, the following detailed description is provided in conjunction with the accompanying drawings and preferred embodiments, based on the specific implementation methods, structures, features and effects of the present invention.

[0021] This application provides a traffic signal control method based on AI edge computing, such as... Figure 1 As shown, the method includes:

[0022] Step S100: Deploy edge nodes at traffic intersections. The edge nodes integrate communication units, TPU accelerators, and multimodal sensors, and can collect traffic flow data in real time and perform local decision-making.

[0023] Preferably, edge nodes are deployed on the roadside at traffic intersections in the form of cabinets or integrated pole-mounted devices. Each edge node integrates a communication unit, a TPU accelerator, and multimodal sensors. The communication unit refers to a hardware unit that supports vehicle-to-everything (V2X) communication protocols such as the PC5 interface in C-V2X or DSRC / IEEE 802.11p, enabling low-latency, high-reliability data exchange with onboard terminals and sending and receiving traffic request messages. This communication unit also includes a cellular network or fiber optic backhaul interface for coordinating data sharing with adjacent edge nodes or a higher-level management platform. The TPU accelerator is a processor used to accelerate the inference and training of artificial intelligence models. The TPU accelerator enables edge nodes to quickly run complex traffic flow prediction models, conflict detection models, and collaborative decision-making models locally to meet real-time requirements.

[0024] Preferably, the multimodal sensor includes multiple complementary sensing devices, specifically including visual sensors and high-resolution cameras for acquiring video streams and detecting vehicle position, speed, vehicle type, lane occupancy status, etc., through computer vision algorithms; radar sensors, millimeter-wave radar or lidar, for accurately measuring vehicle distance, speed, and angle, whose detection is unaffected by lighting or weather, and can provide high-precision target tracking data; other sensing sources may also include V2X data received by roadside units from connected vehicles, such as BSM messages, as a supplement and verification to physical sensing. Edge nodes collect multimodal sensor data and communication messages in real time, fuse them to generate traffic flow data, and perform local computational decisions, thereby ensuring that traffic signal control achieves extremely low response latency, high reliability, and autonomy.

[0025] In step S200, the vehicle terminal collects the vehicle's priority status and driving route, and generates an encrypted passage request message. The passage request message includes at least the vehicle identification, the expected sequence of target intersections to be passed, and the estimated time window for arriving at each intersection.

[0026] Preferably, an in-vehicle terminal installed inside the vehicle, possessing computing, storage, and communication functions—that is, an in-vehicle unit or the central gateway / domain controller of an intelligent connected vehicle—automatically determines and collects the vehicle's priority status based on its attributes or tasks. For example, emergency vehicles such as ambulances, fire trucks, and police cars activate the highest priority when performing tasks; buses and BRT vehicles are assigned high priority; special service vehicles such as emergency rescue vehicles and VIP vehicles may have specific priorities; and ordinary vehicles default to ordinary priority. Simultaneously, the in-vehicle navigation system plans and obtains a driving route in real time based on the destination, real-time traffic conditions, and driver selection, including multiple ordered sequences of roads and intersections. Then, the collected priority status codes and the key intersection ID sequences in the planned route are used as input, combined with the vehicle's... The vehicle's current real-time location, speed, acceleration, and acquired historical / real-time travel time data for road segments are used to predict and calculate the estimated time range for each target intersection on the vehicle's expected path using a kinematic model, thus determining the estimated time window. Then, the vehicle's unique identifier, priority status, target intersection sequence, and corresponding estimated time window are encapsulated into a travel request message according to a predefined communication protocol format. The message is digitally signed and / or encrypted using an asymmetric encryption algorithm to ensure its authenticity, integrity, and confidentiality, preventing forgery or tampering. The vehicle identifier is used to accurately identify the vehicle, such as a vehicle electronic identification number or temporary communication session ID. The expected target intersection sequence clearly identifies the logical identifiers of all intersections the vehicle plans to pass through sequentially.

[0027] In step S300, the detector or RSU device forwards the passage request message to the edge node of the target intersection. After receiving the passage request message, the edge node splits the intersection right-of-way into continuous spatiotemporal resource slots according to the time dimension, wherein each slot defines the lane and time period that can be occupied.

[0028] Step S300 further includes extracting an estimated time window from the traffic demand message to determine the time allocation range of the intersection right-of-way; dividing the time allocation range of the intersection right-of-way into time units according to a preset time granularity; combining each time unit with the lane space range to generate a spatiotemporal resource slot with a unique identifier; and classifying the spatiotemporal resource slots according to the priority marker in the vehicle identification to generate a set of spatiotemporal resource slots.

[0029] Preferably, vehicles cannot directly communicate with the edge nodes of the target intersection. Instead, roadside detectors or RSU devices act as relay nodes, receiving passage request messages from the vehicle-mounted terminals and forwarding them to the corresponding edge nodes via a roadside communication network (such as fiber optic or wireless backhaul links). The detectors identify the vehicle's identity, direction of travel, priority status, and estimated arrival time, uploading the identification results to the edge nodes. The edge nodes perform spatiotemporal resource slot allocation and conflict detection based on the vehicle's passage request information. The edge nodes receive encrypted passage request messages through their communication units, decrypt and verify them, and then extract an estimated time window from the message to determine the time period for resource planning and scheduling for the vehicle at the intersection. This time window can be appropriately extended forward and backward to handle requests and transitions from other vehicles, thus determining the time allocation range for the intersection's right-of-way. This time allocation range is then divided into multiple equal-length time units according to a preset time granularity, where the preset time granularity is the basic time resolution. For example, 100 milliseconds, 200 milliseconds, or 1 second; then, each time unit is bound to a lane space range combination, where the lane space range typically refers to one or more permitted lanes in an approach direction, generating multiple time unit-lane combinations, determining multiple spatiotemporal resource slots, each with a unique identifier containing the code for the permitted lane and time period, where the permitted lane explicitly specifies the physical lane through which the vehicle that obtains the slot can travel, and the time period is the precise start and end time of the right of passage; finally, the generated spatiotemporal resource slots are pre-classified according to the priority marker in the vehicle identity identifier to create a hierarchical resource pool, for example, a high-priority resource slot set, which is some or all of the slots generated within the request time window reserved for high-priority vehicles, with higher allocation weight; a general resource slot set, which is the time slot outside the coverage of priority vehicle demand or slots specially allocated for bidding by ordinary vehicles, ultimately determining a structured set of spatiotemporal resource slots.

[0030] Step S400: Compare the current passage demand messages of all vehicles, detect the occupation conflicts of the same spatiotemporal resource slots, and output the list of conflicting slots and the set of associated vehicles.

[0031] Step S400 further includes parsing the spatiotemporal resource slot requested by each passage demand message; establishing a spatiotemporal resource slot occupancy status table to record the requesting vehicles for each spatiotemporal resource slot; traversing the spatiotemporal resource slot occupancy status table to identify the spatiotemporal resource slots requested by multiple passage demand messages; and recording the spatiotemporal resource slots requested by multiple passage demand messages and their requesting vehicle information as a conflict slot list and an associated vehicle set.

[0032] Preferably, all current vehicle traffic demand messages are parsed in parallel. Based on the target intersection ID, estimated time window, and vehicle intent (e.g., straight ahead or left turn) in the message, it is matched with the locally generated set of spatiotemporal resource slots to determine the specific resource slot requested by the vehicle at this intersection. For example, a left-turning vehicle may request to occupy five consecutive slots of 200 milliseconds each in the "East Entrance Left Turn Lane" within the next 15 to 20 seconds. This generates a list of the specific spatiotemporal resource slots requested by each traffic demand message. Then, a dynamically updated table is created with all generated spatiotemporal resource slots as rows. The occupancy status field for each slot is initially empty or idle. When a vehicle requests to occupy a slot, the vehicle's identifier is added to the record for that slot, thus constructing the spatiotemporal... The resource slot occupancy status table clearly records the requesting vehicles for each spatiotemporal resource slot. For example, if multiple vehicle IDs exist in the requesting vehicle list for a certain slot, it indicates that there is contention for that slot. The spatiotemporal resource slot occupancy status table is then traversed to check the number of entries in the requesting vehicle list. If the list is empty or contains only one vehicle ID, it means there is no conflict, and the slot can be safely allocated. If the list contains two or more vehicle IDs, it indicates that a conflict has been detected, thus identifying the spatiotemporal resource slots requested by multiple passage request messages. Finally, the spatiotemporal resource slots requested by multiple passage request messages and their requesting vehicle information are recorded as a conflict slot list and an associated vehicle set. The conflict slot list contains multiple unique identifiers for the slots, and the associated vehicle set contains the identity identifiers of all vehicles requesting to occupy that slot.

[0033] Step S500: Based on the conflict slot list and associated vehicle set, for the conflicting spatiotemporal resource slots, the edge node initiates a lightweight blockchain auction to generate auction results, which include the winning vehicle and the precisely allocated spatiotemporal slot.

[0034] Preferably, for conflicting spatiotemporal resource slots, edge nodes take the list of conflicting slots, the set of associated vehicles and their respective priority status, available credit points, and real-time energy consumption data as input to initiate a lightweight blockchain auction. This involves calling an auction smart contract deployed on a regional consortium blockchain. Vehicles submit transactions containing credit point bids and comprehensive quotes based on energy consumption weights within the time specified in the contract, according to their own vehicle priority and energy consumption status. The smart contract calculates the final effective weight for each vehicle's bid according to a preset, publicly available weight calculation formula and sorts the effective weights of all bidders. The vehicle with the highest effective weight wins the spatiotemporal slot. The winning vehicle ID, the precise slot ID assigned, and the actual credit points deducted are stored on the blockchain to generate the auction result.

[0035] Preferably, high-priority vehicles are auctioned using credit points. Credit points are pre-allocated non-monetary digital rights certificates based on a vehicle's social function and public service attributes, representing priority passage. A certain amount of credit points is injected based on vehicle type and mission status; for example, an ambulance on duty might be automatically allocated 1000 credit points. When a high-priority vehicle conflicts with other vehicles, it can commit to paying a certain number of credit points to bid for the required spacetime slots, ensuring that emergency vehicles can win in most conflicts by paying their special credit points, thereby achieving priority passage. For ordinary vehicles, energy consumption weighting is used for bidding. Specifically, vehicles use onboard... Sensors collect and report real-time energy consumption data, such as the remaining battery percentage, estimated remaining range, or energy consumption per unit mile. The vehicle's energy consumption data is converted into a bidding weight multiplier according to a preset weight function. The energy consumption bidding weight = base weight + (1 - remaining battery percentage) × sensitivity coefficient. The bidding weight of an electric vehicle with extremely low battery (10%) may be amplified to 1.8, while the weight of a fully charged vehicle (90%) may only be 1.1. The final comprehensive bid of the vehicle is the product of the base bid and the energy consumption bidding weight, thereby introducing energy efficiency and social benefits, tending to help vehicles in energy distress pass through first, and improving the overall resilience of the road network.

[0036] Furthermore, step S500 also includes setting auction parameters according to the degree of conflict of each spatiotemporal resource slot in the conflict slot list; sending bidding requests containing conflict spatiotemporal resource slot information to vehicles in the associated vehicle set; receiving bidding responses submitted by each vehicle according to its priority status; processing the bidding responses according to preset bidding rules to determine the winning vehicle for each conflict spatiotemporal resource slot; and recording the allocation relationship between the winning vehicle and the corresponding spatiotemporal resource slot as the auction result.

[0037] Preferably, before initiating the auction, the edge node analyzes the conflict level of each spatiotemporal resource slot in the conflict slot list, including assessing the number of competing vehicles, the number of vehicles requesting the same slot (three vehicles competing is more competitive than two), the combination of vehicle types, whether it includes high-priority vehicles such as emergency vehicles, and the spatiotemporal criticality of the slot, as well as the slot's position on the timeline. Based on the assessment results, the node dynamically sets the rules and parameters for this auction, which may include the auction type ("open incremental bidding", "sealed bidding", or "composite bidding"), the bidding duration (the bidding time may be shortened for quick decision-making when the conflict level is high or high-priority vehicles are involved), the reserve price or weighting coefficient (setting a higher initial credit point advantage for high-priority vehicles or adjusting the energy consumption weight of their bids based on the vehicle's energy consumption status). As the auction initiator, the edge node sends a structured bidding request message to all vehicles in the associated vehicle set through low-latency V2X communication. The message includes conflicting spatiotemporal resource slot information, auction parameters, and related auction smart contracts.

[0038] Preferably, after receiving a bidding request, each vehicle's onboard terminal generates a bidding response based on its own priority status. High-priority vehicles can use pre-allocated credit points to bid, sending an encrypted message containing their bid commitment and necessary status proof back to the edge node or submitting it to the blockchain smart contract address before the deadline. This process gathers information on the true bidding intentions and capabilities of all competitors. The auction smart contract collects all correctly formatted bidding responses submitted within the specified time, verifies the validity of their digital signatures and status proofs, and processes each bid according to preset bidding rules. For ordinary vehicles, the final valid bid is determined by the auction smart contract. =Base bid × Energy consumption weight coefficient. For high-priority vehicles, the final effective bid = the number of credits committed to be paid. Then, the final effective bids of all bidders for the same conflict slot are compared. Following the principle of "highest bidder wins", the winning vehicle is automatically determined. If there are the same highest bid, the decision is made according to preset rules, such as the order of submission time or starting a second round of sealed bidding. Finally, the binding relationship between the identity of the winning vehicle of each conflict spacetime resource slot and the unique identifier of the assigned spacetime resource slot is written as the auction result into the distributed ledger of the blockchain. Edge nodes and all relevant vehicles can query and verify this auction result.

[0039] Furthermore, step S500 also includes: conducting a first round of open bidding for the conflicting spacetime resource slots using a gradually decreasing price-based open bidding method, and updating the current best bid in real time; when multiple identical best bids are detected, initiating a second round of sealed bidding; determining the winning vehicle based on the second-highest price principle according to the results of the second round of sealed bidding; and writing the binding relationship between the winning vehicle and the corresponding spacetime resource slot into the distributed ledger.

[0040] Preferably, the auction smart contract starts with a preset high starting price and automatically and gradually lowers the current effective bid at fixed time intervals (e.g., every 100 milliseconds) to conduct the first round of open bidding for conflicting spacetime resource slots. Any vehicle wishing to acquire that slot only needs to submit a bid by sending an "accept" signal or triggering an on-chain transaction when the price drops to the expected level. If a vehicle bids, that price is recorded as the "current best bid," the auction clock continues, the price continues to decrease, and the current best bid is updated in real time. At the end of the first round of open bidding, if it is detected that two or more vehicles have bid at the exact same final price, indicating that the open bidding could not distinguish the payment intentions of these vehicles, a second round of sealed bidding is initiated for the vehicles with the same bid. That is, the participating vehicles independently submit higher sealed bids without knowing the bids of other competitors, representing their true highest price that they are willing to pay for the resource. After all sealed bids are submitted, the vehicle with the highest bid is the winner, and payment is made according to the second-highest price principle based on the results of the second round of sealed bidding. Finally, the binding relationship between the winning vehicle, the corresponding spacetime resource slot, and the actual payment bid are written into the distributed ledger, thereby ensuring a balance between auction efficiency and speed.

[0041] In step S600, adjacent edge nodes share the auction results and perform coordinated analysis through distributed machine learning to generate signal control instructions, which include phase timing instructions, phase duration instructions, and phase switching instructions.

[0042] Step S600 further includes: encapsulating the auction results containing the binding relationship between the winning vehicle and the spatiotemporal resource slot into a coordination data packet; sending the coordination data packet to the adjacent edge node through a secure channel; after the adjacent edge node verifies the coordination data packet, extracting the auction result information therein; generating a coordination scheme through a distributed machine learning model based on the auction result information and local traffic data, and encoding the coordination scheme into signal control instructions.

[0043] Preferably, adjacent edge nodes share the auction results. That is, after an edge node at an intersection completes the auction for several conflicting spatiotemporal slots, it encapsulates the auction results, which include the binding relationship between the winning vehicle and the spatiotemporal resource slot, into a coordination data packet. This makes it a complete information unit that can be understood and processed by other nodes. The coordination data packet is sent to adjacent edge nodes through a low-latency secure channel pre-established between edge nodes, ensuring the confidentiality, integrity, and authenticity of the data during transmission. Adjacent edge nodes verify the integrity and authenticity of the coordination data packet, confirming that the source of the coordination data packet is trustworthy and has not been tampered with. After the verification is passed, the auction result information is extracted and stored in the local collaborative decision-making database.

[0044] Preferably, coordinated analysis is performed through distributed machine learning. This involves using auction results received from neighboring nodes as global information and real-time traffic data from the edge nodes themselves as local information. Each edge node deploys the same distributed machine learning model, such as a deep reinforcement learning model or a collaborative prediction model. This model is pre-trained on a large amount of data and may be updated through continuous federated learning. The constraint that "a batch of vehicles will arrive as planned in the future" is used as a strong constraint and inputted into the distributed machine learning model along with the local real-time traffic status. The model calculates the signal timing scheme for the intersection, ensuring that vehicles that have already obtained the right of way at the previous intersection can enjoy a high probability of green light when they arrive at the intersection within the predetermined time window, thus forming a "green wave". At the same time, while meeting the primary objective, the model maximizes the overall traffic efficiency of the intersection, thereby generating a dynamic signal phase and timing optimization scheme, i.e., a coordinated scheme.

[0045] Preferably, the coordination scheme is encoded into standardized signal control instructions, including phase timing instructions, phase duration instructions, and phase switching instructions. This enables the traffic light timing at the intersection to be precisely aligned in time and space with the vehicle communication plan of the previous intersection. Specifically, the phase timing instructions define the order in which the various traffic light colors (red, yellow, and green) appear, generated based on the vehicle turning intentions in the auction results to adjust the logical order of the phases and ensure vehicles travel along the predetermined path. The phase duration instructions define the specific duration of each signal phase, generated based on the precise time windows of the vehicles in the auction results to fine-tune the green light time and ensure the winning vehicle is within its allocated time and space. When a vehicle arrives at its designated lane, it encounters a green light with sufficient duration to allow passage without wasting time. Phase switching instructions define the precise moment for switching from one phase to the next, generated based on real-time collaborative computation results from distributed machine learning to achieve dynamic green waves. For example, an upstream intersection instructs a vehicle to pass at time T1, while a downstream intersection calculates the vehicle's arrival time at its intersection as T2 based on the vehicle's average speed and distance. Consequently, the traffic light is instructed to switch to the phase allowing the vehicle to pass at time T2. This provides a dynamic, reliable, and efficient regional-level passage corridor for high-priority or successfully bid vehicles, while maximizing the overall efficiency of the road network.

[0046] Furthermore, step S600 also includes: each edge node training a local model based on local traffic flow data and received auction results; sharing model parameters through a secure aggregation protocol while protecting the privacy of the original data, aggregating model parameters of multiple edge nodes to generate a regional coordination model; and using the regional coordination model to evaluate different signal control schemes and select the optimal coordination scheme.

[0047] Preferably, local traffic flow data, such as vehicle trajectory, speed, queue length, and occupancy rate, is acquired in real time from each edge node using edge node sensors. Auction results containing vehicle passage reservation plans are obtained from neighboring nodes. Each edge node uses its TPU accelerator locally to train a machine learning model using policy gradient updates in deep reinforcement learning or gradient descent in supervised learning. The model learns to generate signal control strategies that maximize local and regional joint benefits based on the current traffic conditions and known future passage demands, thus determining the local model. During the training process, each edge node only sees its own raw data, thereby protecting data privacy and security.

[0048] Preferably, a secure aggregation protocol is used to share model parameters while protecting the privacy of the original data. This secure aggregation protocol is a cryptographic protocol where each edge node, before sending out its local parameter updates, first homomorphically encrypts them or adds differential privacy noise, and then divides them into secret shares to distribute to multiple participants. The aggregation server collects all encrypted or fragmented parameters, unable to decrypt or restore the original parameters of any single node. It can only calculate the average or weighted sum of all node parameters in encrypted form or through secure multi-party computation, outputting the globally averaged model parameter update. Each edge node merges the aggregated global parameter update into its local model. After multiple rounds of local training-secure aggregation-parameter fusion iterations, the decision logic of each edge node ultimately integrates the traffic patterns, congestion patterns, and collaborative experience of the entire participating aggregation area to determine the regional coordination model. Using the regional coordination model, different signal control schemes are evaluated based on green wave achievement rate, overall traffic efficiency, etc., and the scheme with the highest comprehensive score is selected as the optimal coordination scheme, which is then encoded into executable signal control instructions.

[0049] Furthermore, step S600 also includes establishing an evaluation function for each edge node that includes traffic efficiency indicators, energy consumption cost indicators, and fairness indicators; calculating the evaluation function scores of different signal control schemes based on the vehicle allocation information in the auction results; and obtaining the optimal coordination scheme by iteratively calculating to make the evaluation function of each edge node reach a Nash equilibrium state.

[0050] Preferably, a quantitative satisfaction evaluation function is defined for each edge node, including traffic efficiency indicators, energy consumption cost indicators, and fairness indicators. Among them, the traffic efficiency indicator measures the traffic capacity of the intersection, such as the reciprocal of the total vehicle delay at the intersection, the intersection throughput, and the on-time rate of the winning vehicle in the auction actually receiving a green light at the intersection; the energy consumption cost indicator measures the system energy consumption caused by signal control, such as the reciprocal of the estimated total additional fuel / electricity consumption of all vehicles at the intersection due to stopping, starting, acceleration, and deceleration; the fairness indicator is used to measure the fairness of resource allocation, such as the reciprocal of the variance of waiting time for vehicles in each direction and priority, and the degree to which the maximum waiting time for vehicles in non-priority directions does not exceed a certain threshold; a comprehensive score is generated by weighting the traffic efficiency indicator, energy consumption cost indicator, and fairness indicator, with the weights set according to the importance of the three evaluation indicators.

[0051] Preferably, the vehicle allocation information from the auction results is used as the core input and constraint. When a node proposes a candidate signal control scheme and assumes a set of neighboring schemes, micro-level traffic simulation is performed to ensure that the winning vehicle can pass through its obtained spatiotemporal slot. Then, the traffic flow results such as vehicle trajectories, delays, and number of stops generated by the simulation are substituted into the established evaluation function to calculate the evaluation function score of the intersection under the combination of signal control schemes, thus determining the evaluation function scores of different signal control schemes. Then, through iterative calculation, the evaluation functions of each edge node reach a Nash equilibrium state, that is, no edge node can improve its own evaluation function score by unilaterally changing its own signal control parameters while keeping the signal control schemes of all other nodes unchanged. Finally, the optimal coordination scheme is obtained, thereby ensuring the generation of an executable, stable, and optimally efficient distributed green wave coordination scheme.

[0052] Furthermore, step S600 also includes each edge node proposing initial signal control parameters based on the auction results and local traffic data; exchanging the initial control parameters and evaluation function calculation results of adjacent nodes, and calculating the impact of unilateral adjustment of control parameters on the overall evaluation function; when any edge node cannot improve its own evaluation function by unilaterally adjusting control parameters, it is determined that a Nash equilibrium state has been reached.

[0053] Preferably, each edge node utilizes its local model to generate green light duration, phase difference, and lane occupancy priority as initial signal control parameters based on the auction results and local traffic data such as general traffic flow and queuing conditions. The green light duration is the green light time allocated to different phases (e.g., north-south straight traffic, east-west left turns); the phase difference is the offset time between the opening of each phase at the intersection relative to a reference time point; and the lane occupancy priority is a fine-tuning weight for the right-of-way of different lanes during complex phases or overlapping green light periods. For example, in a phase where left turns and oncoming straight traffic are simultaneously permitted, priority is given to lanes containing the winning vehicle from the auction. A comprehensive score for the initial signal control parameters is calculated using an evaluation function. This reflects the overall benefit of the intersection under the current strategy combination. Then, the initial control parameters of adjacent nodes are exchanged. That is, after receiving the signal control schemes and benefit information of all neighboring nodes, each node fine-tunes one control parameter while keeping the strategies of all neighboring nodes unchanged, and recalculates the overall score using the evaluation function to determine the evaluation function calculation result. Then, the impact of unilateral adjustment of control parameters on the overall evaluation function is compared to obtain the strategy adjustment direction that can improve its own overall benefit. When any edge node cannot improve its own evaluation function by unilaterally adjusting control parameters, it is determined that a Nash equilibrium state has been reached, and the current control parameter scheme is taken as the final optimal coordination scheme.

[0054] In step S700, the edge node uses a random number generator to generate a dynamic key, signs the signal control command, and distributes the signed signal control command to the associated vehicles and intersection traffic light controllers for traffic light switching control.

[0055] Preferably, before sending a command, the edge node calls a cryptographically secure random number generator in its encryption library to generate a temporary dynamic key pair, containing a private key and a corresponding public key, used to sign the signal control command. This pair is discarded after use to enhance the security of command transmission. The dynamic key is used to sign the signal control command; that is, a cryptographic hash value is calculated using SHA-256 to obtain a data fingerprint, and then the dynamic private key is used to encrypt the data fingerprint to generate an additional digital signature. The signed signal control command contains the original signal control command, the digital signature, and the dynamic public key. Then, the signed signal control command is distributed to the associated vehicles and intersection traffic light controllers through a secure channel. After receiving the signed signal control command, the vehicle / signal controller recalculates the hash of the original command data using the attached dynamic public key, decrypts the signature with the public key, compares the two hash values, and performs identity and integrity verification. Once the verification is successful, the vehicle trusts the signal control command, and its assisted driving or onboard system plans its speed according to the signal control command in order to arrive at the intersection and pass smoothly within the specified time window. At the same time, the signal light controller decodes the signal control command and controls the corresponding signal light group to switch to green light during the time period corresponding to the precise time slot, creating passage conditions for the vehicle. This realizes the intelligent allocation and management of scarce time and space resources in complex traffic environments, thereby improving the efficiency of traffic area collaboration and the overall traffic safety and reliability.

[0056] In the above text, refer to Figure 1 This paper describes in detail a traffic signal control method based on AI edge computing according to embodiments of the present invention. Next, reference will be made to... Figure 2 This invention describes a traffic signal control device based on AI edge computing according to an embodiment of the present invention.

[0057] The traffic signal control device based on AI edge computing according to embodiments of the present invention addresses the technical problems in existing traffic signal control, such as the mismatch between fixed timing and actual traffic flow changes, the difficulty in adapting green wave coordination on main roads to dynamic traffic flow changes, latency and fluctuations in cloud control networks, and traffic signal disorder caused by priority vehicle passage. It achieves the technical effects of improving traffic efficiency and real-time reliability of traffic signal control when priority vehicles and ordinary vehicles share the road, enhancing the collaborative control capabilities of adjacent intersections, and reducing the pressure on cloud control. Figure 2 As shown, the traffic signal control device based on AI edge computing includes: an edge node deployment module 10, a passage message generation module 20, a passage message splitting module 30, an occupancy conflict detection module 40, an auction result generation module 50, a coordination analysis module 60, and a signal control module 70.

[0058] The edge node deployment module 10 is used to deploy edge nodes at traffic intersections. Each edge node integrates a communication unit, a TPU accelerator, and a multimodal sensor, enabling real-time acquisition of traffic flow data and execution of local decisions. The passage request message generation module 20 is used by the vehicle terminal to collect the vehicle's priority status and driving path, generating an encrypted passage request message. This message includes at least the vehicle's identification, the expected sequence of target intersections, and the estimated time window for arrival at each intersection. The passage request message splitting module 30 is used by the detector or RSU device to forward the passage request message to the edge nodes at the target intersections. After receiving the message, the edge nodes split the intersection right-of-way into continuous spatiotemporal resource slots based on the time dimension, where each slot defines the allowed lane and time period. The occupancy conflict detection module 40 is used to compare the passage requests of all vehicles currently in operation. The system detects conflicts in the same spatiotemporal resource slots and outputs a list of conflicting slots and a set of associated vehicles. An auction result generation module 50, based on the list of conflicting slots and the set of associated vehicles, initiates a lightweight blockchain auction for conflicting spatiotemporal resource slots at the edge nodes, generating auction results including the winning vehicle and its precisely allocated spatiotemporal slot. A coordination analysis module 60, used by adjacent edge nodes to share the auction results, performs coordination analysis through distributed machine learning, and generates signal control instructions, including phase timing instructions, phase duration instructions, and phase switching instructions. A signal management module 70, used by edge nodes to generate a dynamic key using a random number generator, signs the signal control instructions, and distributes the signed signal control instructions to associated vehicles and intersection traffic light controllers for traffic light switching management.

[0059] The specific configuration of the traffic message splitting module 30 will be described in detail below. The traffic message splitting module 30 further includes: extracting an estimated time window from the traffic demand message to determine the time allocation range of the intersection right-of-way; dividing the time allocation range of the intersection right-of-way into time units according to a preset time granularity; combining each time unit with a lane space range to generate a spatiotemporal resource slot with a unique identifier; and classifying the spatiotemporal resource slots according to the priority marker in the vehicle identification identifier to generate a set of spatiotemporal resource slots.

[0060] The specific configuration of the occupancy conflict detection module 40 will be described in detail below. The occupancy conflict detection module 40 further includes: parsing the spatiotemporal resource slot requested by each passage demand message; establishing a spatiotemporal resource slot occupancy status table to record the requesting vehicles for each spatiotemporal resource slot; traversing the spatiotemporal resource slot occupancy status table to identify spatiotemporal resource slots requested by multiple passage demand messages; and recording the spatiotemporal resource slots requested by multiple passage demand messages and their requesting vehicle information as a conflict slot list and an associated vehicle set.

[0061] The specific configuration of the auction result generation module 50 will be described in detail below. The auction result generation module 50 further includes: setting auction parameters based on the conflict level of each spatiotemporal resource slot in the conflict slot list; sending bidding requests containing conflicting spatiotemporal resource slot information to vehicles in the associated vehicle set; receiving bidding responses submitted by each vehicle according to its priority status; processing the bidding responses according to preset bidding rules to determine the winning vehicle for each conflicting spatiotemporal resource slot; and recording the allocation relationship between the winning vehicle and the corresponding spatiotemporal resource slot as the auction result.

[0062] The specific configuration of the auction result generation module 50 will be described in detail below. The auction result generation module 50 further includes: using a progressively decreasing open bidding method to conduct the first round of open bidding for conflicting spacetime resource slots, updating the current best bid in real time; when multiple identical best bids are detected, initiating a second round of sealed bidding; determining the winning vehicle according to the second-highest price principle based on the results of the second round of sealed bidding; and writing the binding relationship between the winning vehicle and the corresponding spacetime resource slot into the distributed ledger.

[0063] The specific configuration of the coordination analysis module 60 will be described in detail below. The coordination analysis module 60 further includes: encapsulating the auction results, which contain the binding relationship between the winning vehicle and the spatiotemporal resource slot, into a coordination data packet; sending the coordination data packet to adjacent edge nodes through a secure channel; after verifying the coordination data packet, the adjacent edge nodes extract the auction result information from it; based on the auction result information and local traffic data, generating a coordination scheme through a distributed machine learning model, and encoding the coordination scheme into signal control instructions.

[0064] The specific configuration of the coordination analysis module 60 will be described in detail below. The coordination analysis module 60 further includes: each edge node training a local model based on local traffic flow data and received auction results; sharing model parameters through a secure aggregation protocol while protecting the privacy of the original data, aggregating model parameters from multiple edge nodes to generate a regional coordination model; and using the regional coordination model to evaluate different signal control schemes and select the optimal coordination scheme.

[0065] The specific configuration of the coordination analysis module 60 will be described in detail below. The coordination analysis module 60 further includes: establishing an evaluation function for each edge node that includes traffic efficiency indicators, energy consumption cost indicators, and fairness indicators; calculating the evaluation function scores of different signal control schemes based on the vehicle allocation information in the auction results; and obtaining the optimal coordination scheme by iteratively calculating to bring the evaluation functions of each edge node to a Nash equilibrium state.

[0066] The specific configuration of the coordination analysis module 60 will be described in detail below. The coordination analysis module 60 further includes: each edge node proposing initial signal control parameters based on the auction results and local traffic data; exchanging the initial control parameters and evaluation function calculation results of adjacent nodes, and calculating the impact of unilateral adjustment of control parameters on the overall evaluation function; when any edge node cannot improve its own evaluation function by unilaterally adjusting control parameters, it is determined that a Nash equilibrium state has been reached.

[0067] The traffic signal control device based on AI edge computing provided in this embodiment of the invention can execute the traffic signal control method based on AI edge computing provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method execution.

[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A traffic signal control method based on AI edge computing, characterized in that, The method includes: Edge nodes are deployed at traffic intersections. These edge nodes integrate communication units, TPU accelerators, and multimodal sensors, enabling them to collect traffic flow data in real time and execute local decisions. The vehicle terminal collects the vehicle's priority status and driving route, and generates an encrypted passage request message. The passage request message includes at least the vehicle identification, the expected sequence of target intersections to be passed, and the estimated time window for reaching each intersection. The detector or RSU device forwards the passage request message to the edge node of the target intersection. After receiving the passage request message, the edge node splits the intersection right-of-way into continuous spatiotemporal resource slots according to the time dimension. Each slot defines the lane and time period that can be occupied. Compare the current passage demand messages of all vehicles, detect the occupation conflicts of the same spatiotemporal resource slots, and output the list of conflicting slots and the set of associated vehicles; Based on the list of conflicting slots and the associated vehicle set, for conflicting spatiotemporal resource slots, the edge node initiates a lightweight blockchain auction to generate auction results, including: Auction parameters are set according to the degree of conflict of each spatiotemporal resource slot in the conflict slot list; Send bidding requests containing conflicting spatiotemporal resource slot information to vehicles in the associated vehicle set; Receive bidding responses from each vehicle based on its priority status; Bidding responses are processed according to preset bidding rules to determine the winning vehicle for each conflicting spatiotemporal resource slot, including: A public bidding process with progressively decreasing prices is adopted to conduct the first round of public bidding for conflict spacetime resource slots, and the current best bid is updated in real time. When multiple identical best bids are detected, a second round of sealed bidding is initiated. Based on the results of the second round of sealed bidding, the winning vehicle is determined according to the principle of the second highest price, and the binding relationship between the winning vehicle and the corresponding spatiotemporal resource slot is written into the distributed ledger; The allocation relationship between the winning vehicle and the corresponding spatiotemporal resource slot is recorded as the auction result, which includes the winning vehicle and the precise spatiotemporal slot allocated. Adjacent edge nodes share the auction results, coordinate and analyze them through distributed machine learning, and generate signal control instructions, including: The auction results, which include the binding relationships between the winning vehicles and the spacetime resource slots, are encapsulated into a coordination data packet; Coordination data packets are sent to adjacent edge nodes via a secure channel; After the adjacent edge nodes verify and coordinate the data packet, they extract the auction result information from it. Based on the auction results and local traffic data, a coordination scheme is generated through a distributed machine learning model, and the coordination scheme is encoded into signal control instructions, which include phase timing instructions, phase duration instructions, and phase switching instructions. Edge nodes use a random number generator to generate a dynamic key, sign the signal control command, and distribute the signed signal control command to associated vehicles and intersection traffic light controllers for traffic light switching control.

2. The traffic signal control method based on AI edge computing as described in claim 1, characterized in that, After receiving the passage request message, the edge node divides the intersection right-of-way into continuous spatiotemporal resource slots according to the time dimension. Each slot defines the lane and time period that can be occupied, including: Extract the estimated time window from the traffic demand message to determine the time allocation range of the right-of-way at the intersection; The time allocation range of the right-of-way at the intersection is divided into time units according to a preset time granularity; Each time unit is combined with the lane space range to generate a spatiotemporal resource slot with a unique identifier; Based on the priority markers in the vehicle identification, the spatiotemporal resource slots are classified to generate a set of spatiotemporal resource slots.

3. The traffic signal control method based on AI edge computing as described in claim 1, characterized in that, Compare the current passage demand messages of all vehicles, detect conflicts in the occupancy of the same spatiotemporal resource slots, and output a list of conflicting slots and a set of associated vehicles, including: Parse the spatiotemporal resource slots requested in each passage requirement message; Establish a spatiotemporal resource slot occupancy status table to record the requesting vehicles for each spatiotemporal resource slot; Traverse the spatiotemporal resource slot occupancy status table to identify the spatiotemporal resource slots requested by multiple passage demand messages; The spatiotemporal resource slots requested by multiple passage demand messages and their requesting vehicle information are recorded as a conflict slot list and an associated vehicle set.

4. The traffic signal control method based on AI edge computing as described in claim 1, characterized in that, Based on the auction results and local traffic data, a coordination scheme is generated through distributed machine learning, including: Each edge node trains a local model based on local traffic flow data and the received auction results; Model parameters are shared through a secure aggregation protocol while protecting the privacy of the original data. By aggregating the model parameters of multiple edge nodes, a regional coordination model is generated. The regional coordination model is used to evaluate different signal control schemes and select the optimal coordination scheme.

5. The traffic signal control method based on AI edge computing as described in claim 4, characterized in that, The regional coordination model is used to evaluate different signal control schemes and select the optimal coordination scheme, including: Establish an evaluation function for each edge node that includes traffic efficiency indicators, energy consumption cost indicators, and fairness indicators; Based on the vehicle allocation information in the auction results, the evaluation function scores of different signal control schemes are calculated; By iteratively calculating, the evaluation functions of each edge node are brought to a Nash equilibrium state, thus obtaining the optimal coordination scheme.

6. The traffic signal control method based on AI edge computing as described in claim 5, characterized in that, The evaluation function of each edge node is brought to Nash equilibrium through iterative calculation, including: Each edge node proposes initial signal control parameters based on the auction results and local traffic data; Exchange the initial control parameters and evaluation function calculation results of adjacent nodes, and calculate the impact of unilateral adjustment of control parameters on the overall evaluation function; When any edge node fails to improve its evaluation function by unilaterally adjusting its control parameters, it is determined that a Nash equilibrium state has been reached.

7. A traffic signal control device based on AI edge computing, characterized in that, The apparatus is used to implement the traffic signal control method based on AI edge computing as described in any one of claims 1 to 6, and the apparatus comprises: An edge node deployment module is used to deploy edge nodes at traffic intersections. The edge nodes integrate communication units, TPU accelerators, and multimodal sensors, and can collect traffic flow data in real time and perform local decisions. The passage message generation module is used by the vehicle terminal to collect the vehicle's priority status and driving route, and generate an encrypted passage request message. The passage request message includes at least the vehicle identification, the expected sequence of target intersections to be passed, and the estimated time window for arriving at each intersection. The passage message splitting module is used to forward the passage request message to the edge node of the target intersection by the detector or RSU device. After receiving the passage request message, the edge node splits the right-of-way of the intersection into continuous spatiotemporal resource slots according to the time dimension. Each slot defines the lane and time period that can be occupied. The occupancy conflict detection module is used to compare the passage demand messages of all vehicles at present, detect occupancy conflicts of the same spatiotemporal resource slots, and output a list of conflicting slots and a set of associated vehicles. The auction result generation module is used to initiate a lightweight blockchain auction on edge nodes for conflicting spatiotemporal resource slots based on the conflicting slot list and associated vehicle set, and generate auction results, including: Auction parameters are set according to the degree of conflict of each spatiotemporal resource slot in the conflict slot list; Send bidding requests containing conflicting spatiotemporal resource slot information to vehicles in the associated vehicle set; Receive bidding responses from each vehicle based on its priority status; Bidding responses are processed according to preset bidding rules to determine the winning vehicle for each conflicting spatiotemporal resource slot, including: A public bidding process with progressively decreasing prices is adopted to conduct the first round of public bidding for conflict spacetime resource slots, and the current best bid is updated in real time. When multiple identical best bids are detected, a second round of sealed bidding is initiated. Based on the results of the second round of sealed bidding, the winning vehicle is determined according to the principle of the second highest price, and the binding relationship between the winning vehicle and the corresponding spatiotemporal resource slot is written into the distributed ledger; The allocation relationship between the winning vehicle and the corresponding spatiotemporal resource slot is recorded as the auction result, which includes the winning vehicle and the precise spatiotemporal slot allocated. The coordination analysis module is used by adjacent edge nodes to share the auction results. It performs coordination analysis through distributed machine learning and generates signal control instructions, including: The auction results, which include the binding relationships between the winning vehicles and the spacetime resource slots, are encapsulated into a coordination data packet; Coordination data packets are sent to adjacent edge nodes via a secure channel; After the adjacent edge nodes verify and coordinate the data packet, they extract the auction result information from it. Based on the auction results and local traffic data, a coordination scheme is generated through a distributed machine learning model, and the coordination scheme is encoded into signal control instructions, which include phase timing instructions, phase duration instructions, and phase switching instructions. The signal control module is used by edge nodes to generate dynamic keys using a random number generator, sign the signal control commands, and distribute the signed signal control commands to associated vehicles and intersection traffic light controllers for traffic light switching control.

Images