Intelligent signal lamp phase optimization method and system based on vehicle infrastructure cooperation

Abstract

The invention discloses an intelligent signal lamp phase optimization method and system based on vehicle infrastructure cooperation. According to the invention, through the dynamic game model, the system can sense multi-dimensional states such as traffic flow and pedestrian waiting time in real time, dynamically adjust demand weights of vehicles, pedestrians and roadside facilities, flexibly allocate green light duration, i.e., multi-direction multi-passing time of vehicles and multi-pedestrian time period to preferentially guarantee the street crossing demand; the average waiting time of vehicles and pedestrians is effectively shortened, the overall traffic efficiency of the intersection is obviously improved, and the suitability of daily traffic operation is obviously improved. The dynamic game model rapidly recognizes special requirements through a state value function, dynamically improves the weight priority of emergency vehicles or pedestrians, guarantees the efficient passing of key scenes (such as the rapid passing of emergency vehicles), avoids the long-time overstock of vehicles in other directions, and improves the safety of emergency vehicles or pedestrians. And the response capability and coordination level of the system to complex and emergent scenes are obviously enhanced.

Description

Technical Field

[0001] The present invention belongs to the field of traffic control technology, and specifically relates to an intelligent traffic light phase optimization method and system based on vehicle-road collaboration. Background Art

[0002] Traffic light phases refer to the sequence and duration of traffic light allocations to traffic flows in different directions during different time periods. They are a core element of urban traffic control, aiming to improve road efficiency, reduce traffic conflicts, and ensure pedestrian and vehicle safety through scientific scheduling and signal timing. A complete traffic light phase typically includes multiple stages, such as green lights for going straight, turning left, and pedestrian crossings. Each stage dynamically regulates traffic flow by setting different green, yellow, and red light times. The design of signal phases must comprehensively consider factors such as intersection traffic volume, number of lanes, traffic composition, and pedestrian crossing needs. Intelligent management is typically achieved through sensor-based control, timing control, or adaptive control systems. Appropriate signal light phases not only alleviate traffic congestion but also reduce the incidence of traffic accidents, improving the overall operational efficiency of urban transportation systems and the overall travel experience.

[0003] However, existing technologies mainly rely on fixed phase durations to control traffic lights, making it difficult to dynamically adjust green light allocation based on real-time traffic flow, pedestrian waiting time, etc., and are prone to problems such as insufficient green lights in the main direction during peak hours leading to vehicle backlogs, and empty green lights during off-peak hours causing waste of resources. At the same time, there is a lag in response to special scenarios (such as the passage of emergency vehicles and concentrated pedestrian crossings), and it is impossible to quickly identify and prioritize key needs, which often leads to delays in emergency vehicles or pedestrians passing illegally due to waiting too long. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent traffic light phase optimization method and system based on vehicle-road collaboration in order to solve the above-mentioned problems.

[0005] The technical solution adopted by the present invention is as follows: an intelligent traffic light phase optimization method based on vehicle-road collaboration, the method comprising the following steps: S1: Through roadside sensors, on-board equipment, and the traffic cloud platform, data such as vehicle location, speed, type, and pedestrian needs are synchronously acquired and transmitted to the regional control center in real time, providing raw materials for subsequent fusion; S2: Based on the multi-source data collected by S1, the timestamps and spatial coordinates are calibrated, and federated learning is used to fuse the local vehicle-side data and the global roadside data to generate a full-factor situation map of the intersection, providing a unified data source for status assessment; S3: Based on the situation diagram of S2, the traffic operation index is calculated using the fuzzy evaluation method from the four dimensions of flow, efficiency, fairness, and safety to identify the current phase shortcomings and provide a current basis for demand forecasting; S4: Combining the TPI from S3 with historical data, LSTM is used to predict traffic flow in all directions, pedestrian crossing times, and special vehicle needs in the next five minutes. Connected vehicles are used to obtain destination information to improve accuracy and provide input for the game model. S5: Based on the needs predicted in S4, a three-party game model of "vehicle-road-pedestrian" is constructed. With the goal of minimizing overall delay, reinforcement learning is used to dynamically adjust the weights of each party and output the phase adjustment tendency to provide a decision-making basis for solution generation. S6: Phase plan generation and screening: Based on the adjustment tendency of S5, call the preset phase template and combine it with the predicted data of S4 to generate 3 to 5 candidate plans. Use TPI pre-simulation to select the best 1 to 2 as the plans to be implemented; S7: Collaborative verification and dynamic deviation correction: Send the S6 execution plan to the connected vehicle and pedestrian screens to collect actual driving and street crossing feedback data; if the deviation exceeds 15%, trigger the S5 model to readjust the weight or insert a temporary phase; S8: Full-cycle model iteration; record the execution effect of S7, use transfer learning to update the parameters of the S5 game model and S4 prediction model, and synchronize them to similar intersections in the area to form a "collection-optimization-sharing" collaborative evolution machine.

[0006] In a preferred embodiment, in step S1, multi-source data is collected in real time through the collaboration of roadside sensing equipment, vehicle-mounted terminals and traffic cloud platforms; cameras deployed on the roadside collect image data at a frequency of 25 frames per second, covering a range of 500 meters at each entrance to the intersection; millimeter-wave radar obtains vehicle position and speed information at a scanning frequency of 0.1 seconds per time, and the ranging accuracy is controlled at ±0.1 meters; the on-board OBU equipment uploads vehicle dynamic data every 2 seconds through V2X communication; pedestrian crossing needs are detected in real time by roadside infrared sensors, and the number of waiting people and detention time are updated every 5 seconds.

[0007] In a preferred embodiment, in step S2, the spatiotemporal alignment and fusion of multi-source data are based on timestamp calibration and spatial coordinate unification; time calibration uses the NTP protocol to synchronize the clocks of all devices, and the error is controlled within ±50 milliseconds; the spatial coordinates use the center point of the intersection as the origin to establish a local coordinate system, and the relative coordinates of the roadside sensors and the GPS absolute coordinates of the on-board equipment are converted into a unified coordinate system; the fusion process adopts a federated learning framework, the vehicle-side equipment locally extracts features of the perceived local traffic data, and the roadside center aggregates all vehicle-side features and roadside global data, and updates the global model parameters every 5 minutes. The number of local iterations is set to 3 times, and finally a full-factor traffic situation map covering a 500-meter range of the intersection with an update frequency of 1 second / time is generated.

[0008] In a preferred embodiment, in step S3, the multi-dimensional evaluation of real-time traffic status is based on the integrated full-factor situation map, and a quantitative analysis is carried out from four dimensions: flow, efficiency, fairness, and safety; the flow dimension uses a 5-minute sliding window to count the number of vehicles in each import channel and calculate the real-time flow; the efficiency dimension calculates the average delay time by comparing the actual vehicle passing time with the free flow passing time.

[0009] In a preferred embodiment, in step S4, the dynamic prediction of multi-scenario traffic demand uses an LSTM neural network as the core model, with the input being the fused data of the past 15 minutes, and the hidden layer setting 64 nodes to predict the traffic status in the next 5 minutes; the model training data covers four scenarios: morning rush hour, evening rush hour, off-peak and special events on weekdays, the training cycle is 30 days, and the prediction error is controlled within the range of vehicle flow ±10% and pedestrian waiting time ±5 seconds.

[0010] In a preferred embodiment, in step S5, a data-driven intelligent algorithm is used to find the optimal balance point among the conflicting demands of the three types of traffic participants: vehicle, road, and pedestrian. First, based on the traffic demand predicted in S4 for the next five minutes, the types of entities that currently need to be coordinated are identified. Then, with the overall goal of "minimizing overall delay," connected vehicles, roadside facilities, and pedestrians are defined as the three game entities, each corresponding to a set of dynamically changing strategies. Finally, a reinforcement learning algorithm is used to dynamically adjust the weight distribution of the three parties in the overall goal, ultimately outputting the most reasonable phase adjustment tendency at the current time. The policy gradient algorithm in deep reinforcement learning is used. The state space of the model is composed of the full-factor situation map generated by S2 and the 5-minute demand data predicted by S4. The action space corresponds to the executable phase adjustment operations. The reward function is directly related to the overall goal - by calculating the overall delay reduction after executing a certain action in real time, the model is driven to continuously optimize the strategy. The calculation formula for dynamic weight allocation is: ; Among them, w i (t) represents the weight coefficient of the i-th class subject at time t, α and β are normalization coefficients, base i is the base weight of the subject, scene i (t) is the scenario correction term; this formula, through a two-layer design of "basic weight + scenario correction", ensures stability in daily scenarios while enabling rapid response to special events; The calculation formula of the state value function is: ; Among them, s represents the current traffic status, γ is the discount factor, R i(s) is the immediate benefit of the i-th subject in state s; this function directly links the weights of the three parties with their respective benefits, and the model selects the optimal action by maximizing V(s); this design enables the model decision to reflect the differences in needs of different subjects while always focusing on the overall goal of "minimizing overall delay."

[0011] In a preferred embodiment, in step S6, the intelligent generation of phase schemes and the screening of candidate sets start with a preset basic phase template, and the template is divided according to time period characteristics: the morning peak template has a green light of 90 seconds in the main direction and 45 seconds in the secondary direction, the flat peak template has a green light of 70 seconds in the main direction and 35 seconds in the secondary direction, and the special event template has a green light of 120 seconds in the main direction and 30 seconds in the secondary direction; based on the adjustment tendency output by the game model, the green light duration of the basic template is fine-tuned by ±10 seconds to generate 3 to 5 candidate schemes.

[0012] In a preferred embodiment, in step S7, the real-time verification and dynamic correction of vehicle-road collaboration sends the plan to be executed to the connected vehicle and pedestrian prompt screen through the roadside RSU device, with a sending frequency of 2 seconds per time; the actual driving trajectory of the connected vehicle and the pedestrian crossing behavior are synchronously collected, and the deviation rate between the actual data and the predicted value is calculated.

[0013] In a preferred embodiment, in step S8, the full-cycle closed-loop optimization and model iteration use the full data of 24 hours a day as the training set, and use the transfer learning algorithm to update the game model and prediction model parameters; the source domain of the transfer learning is the current intersection data, and the target domain is similar intersections in the area, and the model parameters are updated once a week.

[0014] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. In the present invention, through a dynamic game model, the system can perceive multi-dimensional states such as traffic flow and pedestrian waiting time in real time, dynamically adjust the demand weights of vehicles, pedestrians, and roadside facilities, and flexibly allocate green light duration - more time is given to directions with more vehicles, and priority is given to crossing needs during periods with more people, effectively shortening the average waiting time of vehicles and pedestrians, significantly improving the overall traffic efficiency of the intersection and significantly improving the adaptability of daily traffic operations.

[0015] 2. In the present invention, the dynamic game model quickly identifies special needs through the state-value function and dynamically increases the weight priority of emergency vehicles or pedestrians. For example, when an ambulance is detected approaching, the green light in the corresponding direction is extended first, and when pedestrians are observed waiting in a concentrated manner, a temporary crossing phase is inserted. This not only ensures efficient passage in key scenarios (such as the rapid passage of emergency vehicles), but also avoids long-term backlogs of vehicles in other directions, significantly enhancing the system's response capability and coordination level to complex and sudden scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the process principle of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0018] Example: Reference Figure 1 , An intelligent traffic light phase optimization method based on vehicle-road collaboration is characterized in that the method comprises the following steps: S1: Real-time collection of multi-source heterogeneous data. Using roadside sensors (cameras, radar), on-board OBUs, and the traffic cloud platform, we simultaneously acquire multi-dimensional data such as vehicle location, speed, type (including buses and emergency vehicles), pedestrian crossing requirements, and historical traffic flow in all directions at the intersection. This data is then transmitted in real time to the regional traffic control center, providing raw material for subsequent fusion processing.

[0019] S2: Spatiotemporal alignment and fusion of multi-source data. Based on the multi-source data collected in S1, time stamp calibration and spatial coordinate unification (for example, establishing a coordinate system with the intersection center as the origin) are performed to eliminate time delays and spatial deviations between different devices. A federated learning algorithm is used to fuse local traffic status perceived by the vehicle with global roadside perception data to form a comprehensive traffic situation map covering a 500-meter radius around the intersection, providing a unified data source for status assessment in S3.

[0020] S3: Multi-dimensional assessment of real-time traffic status. Based on the integrated full-factor situation map from S2, the traffic performance index (TPI) of the current intersection is calculated using a fuzzy comprehensive evaluation method based on four dimensions: flow (real-time traffic flow in each direction), efficiency (average queue length and delay time), fairness (percentage of waiting time for buses / emergency vehicles), and safety (risk level of pedestrian-vehicle conflict). This method identifies shortcomings of the current phase plan (e.g., excessively long queues in a certain direction or pedestrian waiting times exceeding a threshold), providing a current basis for demand forecasting in S4.

[0021] S4: Dynamic Prediction of Multi-Scenario Traffic Demand. Combining the current TPI assessed in S3 with historical traffic data for the same period (same time period, weather conditions, and event types), an LSTM neural network is used to predict vehicle arrival rates at each entrance, pedestrian crossing times, and priority for buses and emergency vehicles within the next five minutes. Specifically for connected vehicles in a vehicle-infrastructure collaborative environment, destination information is obtained through V2X communication to predict whether they will turn at the current intersection, improving prediction accuracy and providing input variables for the game model in S5.

[0022] S5: Construction of a multi-party dynamic game model (a core creative step). To address the multi-scenario requirements predicted in S4, a dynamic game model encompassing the three parties of "vehicle-road-pedestrian" is constructed. With the overall goal of minimizing overall delay, connected vehicles (pursuing their own traffic efficiency), roadside facilities (ensuring overall traffic flow), and pedestrians (ensuring safe and timely crossings) are the game's main players. A reinforcement learning algorithm dynamically adjusts the weights of the three parties (e.g., increasing the weight of emergency vehicles to 0.7 upon detection, and maintaining a minimum weight of 0.3 for pedestrians during normal hours). Combined with the full-factor situation map from S2, the strategy set for each party is updated in real time (e.g., whether to extend the green light in a certain direction or insert a pedestrian crossing phase). The model outputs the current optimal phase adjustment trend, providing a basis for decision-making in S6's solution generation.

[0023] S6: Intelligent generation of phase plans and candidate set screening. Based on the adjustment trends output by the S5 model, the system calls preset basic phase templates (such as the multi-directional balancing template for the morning peak and the single-main direction priority template for the off-peak period) and combines them with the 5-minute demand data predicted by S4 to generate 3-5 candidate phase plans (such as Plan A, which prioritizes buses, Plan B, which reduces pedestrian waiting times, and Plan C, which balances vehicle and passenger demand). The system then pre-simulates the candidate plans using the TPI evaluation metric from S3, selecting one or two plans with the most significant TPI improvements as candidates for execution, providing test subjects for real-time verification in S7.

[0024] S7: Real-time verification and dynamic deviation correction for vehicle-road collaboration. The execution plan selected in S6 is sent to connected vehicles (displaying the estimated time of crossing) and pedestrian screens (displaying the waiting time for crossing) via roadside RSUs. Feedback data is simultaneously collected, including the vehicle's actual driving trajectory (whether it follows the predicted path) and pedestrians' actual crossing behavior (whether they cross early or late). If the actual behavior deviates by more than 15% from the S4 prediction (e.g., a sudden increase in left-turn requests from connected vehicles in a certain direction), the game model in S5 is triggered to recalculate weights and dynamically adjust the green light duration of the phase plan or insert a temporary phase (e.g., adding a 5-second left-turn transition phase) to ensure the plan matches the actual traffic flow.

[0025] S8: Full-cycle closed-loop optimization and model iteration. The solution execution results (TPI improvement, satisfaction feedback from various stakeholders) from the S7 verification process are recorded. Combined with the full data collected in S1, a transfer learning algorithm is used to update the weight parameter library of the S5 game model (e.g., if a pedestrian weight increase significantly improves TPI in a certain scenario, the weight is fixed). This update also applies to the training data of the S4 prediction model. Simultaneously, the optimized model parameters are synchronized with other similar intersections in the region (e.g., intersections with similar main roads), forming a collaborative evolutionary mechanism of "data collection - model optimization - solution sharing" to continuously improve the adaptability of the overall system.

[0026] In step S1, multi-source data is collected in real time through the collaborative efforts of roadside sensing devices, on-board terminals, and the traffic cloud platform. Roadside cameras capture image data at a rate of 25 frames per second, covering a 500-meter radius around each intersection entrance. Millimeter-wave radar acquires vehicle position and speed information at a scanning rate of 0.1 seconds, with ranging accuracy controlled to ±0.1 meters. On-board OBUs upload vehicle dynamic data (including location, speed, type, and destination) every two seconds via V2X communication. Pedestrian crossing requests are detected in real time by roadside infrared sensors, with the number of people waiting and delay time updated every five seconds. All data is transmitted to the regional traffic control center via the 5G network with a latency of ≤200 milliseconds, forming a raw dataset encompassing over 20 dimensions, including vehicle trajectories, pedestrian density, and historical traffic flow. This provides high-frequency, multi-dimensional foundational data for subsequent fusion processing.

[0027] In step S2, the spatiotemporal alignment and fusion of multi-source data is based on timestamp calibration and spatial coordinate unification. Time calibration uses the NTP protocol to synchronize all device clocks, with an error within ±50 milliseconds. A local coordinate system is established with the intersection center as the origin. The relative coordinates of the roadside sensors and the absolute GPS coordinates of the onboard devices (with an accuracy of ±0.5 meters) are converted into a unified coordinate system. The fusion process utilizes a federated learning framework. The vehicle-side devices extract features from local traffic data (such as vehicle behavior within 50 meters). The roadside center aggregates all vehicle-side features with global roadside data (such as overall traffic flow within 200 meters). Global model parameters are updated every five minutes, with a local iteration count of three. Ultimately, a full-factor traffic situation map covering a 500-meter radius of the intersection is generated with an update frequency of one second. This map contains 15 fused dimensions, including traffic flow in all directions, pedestrian waiting times, and emergency vehicle locations.

[0028] In step S3, a multi-dimensional assessment of real-time traffic status is conducted based on the integrated full-factor situation map, conducting a quantitative analysis across four dimensions: flow rate, efficiency, fairness, and safety. In the flow dimension, a 5-minute sliding window is used to count the number of vehicles at each entrance channel and calculate real-time flow rate (vehicles / hour). In the efficiency dimension, the average delay time is calculated by comparing the actual vehicle transit time with the free-flow transit time (theoretical transit time without interference). In the fairness dimension, the proportion of bus waiting time to the total vehicle waiting time is calculated, and pedestrians are monitored to see whether their continuous waiting time exceeds a threshold of 30 seconds. In the safety dimension, the minimum spatial distance between pedestrians and vehicles is calculated (the safety threshold is set at 1.5 meters) to assess conflict risk. Data from each dimension is input into a fuzzy comprehensive evaluation model. The membership function for flow rate is set as a linear mapping from 0 to 1 corresponding to a range of 0 to 800 vehicles / hour, and the membership function for delay time is exponentially decaying (a membership of ≥0.8 for delays ≤ 60 seconds). The final output is a Traffic Performance Index (TPI) ranging from 0 to 100. Lower scores indicate worsening traffic conditions, providing a quantitative basis for demand forecasting.

[0029] In step S4, the multi-scenario dynamic traffic demand forecasting model uses an LSTM neural network as its core model. Input is fused data from the past 15 minutes (including eight-dimensional features such as traffic volume, delay time, and pedestrian waiting time). The hidden layer has 64 nodes, and the model predicts traffic conditions for the next five minutes. The model training data covers four scenarios: weekday morning rush hour (7:00-9:00 AM), evening rush hour (5:00-7:00 PM), off-peak hours (other times), and special events (such as rainfall and large-scale events). The training cycle is 30 days, and the prediction error is kept within ±10% for traffic volume and ±5 seconds for pedestrian waiting time. For connected vehicles (with a penetration rate of ≥30%), the model uses their uploaded destination information to predict turning needs (e.g., an eastbound destination is marked as a left turn), improving prediction accuracy to over 85%. Output forecasts include vehicle arrival rates at each entrance lane, pedestrian crossing times, and bus / emergency vehicle priority demand, providing dynamic input to the game theory model.

[0030] In step S5, a data-driven intelligent algorithm seeks the optimal balance among the conflicting demands of the three traffic participants: vehicles, roads, and pedestrians. The specific process can be summarized as follows: First, based on the traffic demand predicted in step S4 for the next five minutes (e.g., the expected arrival of three ambulances in a certain direction and the time when pedestrians will be concentrated crossing the street), the types of entities requiring coordination are identified. Then, with the overall goal of minimizing overall delay, the three game-playing entities are defined: connected vehicles (pursuing their own traffic efficiency), roadside infrastructure (ensuring overall smoothness), and pedestrians (ensuring safety and timeliness). Each entity is assigned a set of dynamically changing strategies (e.g., vehicles can choose to "accelerate through" or "slow down and wait," roadside infrastructure can choose to "extend the green light" or "switch phases," and pedestrians can choose to "cross the street early" or "wait for a prompt"). Finally, a reinforcement learning algorithm dynamically adjusts the weights of the three parties in the overall goal (e.g., increasing the weight of emergency vehicles when detected and prioritizing pedestrian needs during off-peak hours). The algorithm ultimately outputs the most appropriate phase adjustment preference (e.g., "prioritize extending the green light at the east entrance by 30 seconds" or "inserting a 5-second pedestrian crossing phase").

[0031] In terms of specific algorithm implementation, the model adopts the policy gradient algorithm in deep reinforcement learning. The state space of the model is composed of the full-factor situation map generated by S2 (including 12-dimensional features such as traffic flow in each direction, pedestrian waiting time, and emergency vehicle position) and the 5-minute demand data predicted by S4 (such as "15 new left-turning vehicles will be added to the north entrance in the next 2 minutes"); the action space corresponds to executable phase adjustment operations (such as "green light in the main direction + 10 seconds" and "insert pedestrian phase" and other discrete actions); the reward function is directly related to the overall goal - by calculating in real time the overall delay reduction after executing a certain action (such as the total delay of the original plan is 120 seconds, and the new plan is reduced to 80 seconds, then the reward value is 40), driving the model to continuously optimize the strategy. This data-driven dynamic game mechanism breaks through the limitations of traditional fixed weight distribution, allowing the phase plan to be flexibly adjusted according to real-time scenarios; The calculation formula for dynamic weight allocation is: ; Among them, w i (t) represents the weight coefficient of the i-th subject (i = car / road / person) at time t, α and β are normalization coefficients (to ensure that the sum of the three weights is 1), base i is the basic weight of the subject (e.g. pedestrian is 0.3, vehicle is 0.5, roadside is 0.2 during daily period), scene i (t) is the scene correction item (e.g., when an emergency vehicle is detected, the scene i (t) Improved by 0.4, pedestrian scene i(t) decreases by 0.1). This formula, through a two-layer design of "basic weight + scenario correction", ensures stability in daily scenarios while enabling rapid response to special events (such as ambulances and large-scale events); The calculation formula of the state value function is: ; Among them, s represents the current traffic state (composed of the situation map and predicted data), γ is the discount factor, and R i (s) is the immediate benefit of the i-th agent in state s (e.g., the benefit to vehicles is "time saved by passing through the intersection," the benefit to pedestrians is "reduced waiting time for crossing the street," and the benefit to the roadside is "global delay reduction"). This function directly links the weights of the three parties to their respective benefits. The model selects the optimal action by maximizing V(s) (e.g., if an action results in a benefit of +15 for vehicles, +10 for pedestrians, and +20 for the roadside, and the vehicle weight is higher, then that action is preferred). This design allows the model's decisions to reflect the diverse needs of different agents while remaining focused on the overall goal of minimizing overall delay. In step S6, intelligent phase scheme generation and candidate set screening begin with a pre-set basic phase template. These templates are divided according to time period characteristics: the morning peak template has a green light duration of 90 seconds in the primary direction and 45 seconds in the secondary direction; the off-peak template has a green light duration of 70 seconds in the primary direction and 35 seconds in the secondary direction; and the special event template (such as emergency vehicle passage) has a green light duration of 120 seconds in the primary direction and 30 seconds in the secondary direction. Based on the adjustment trends output by the game model (such as "pedestrian weighting" or "emergency vehicle priority"), the green light duration of the basic template is fine-tuned by ±10 seconds to generate 3-5 candidate schemes (e.g., a green light duration of +10 seconds in the primary direction and -5 seconds in the secondary direction). The candidate schemes are pre-simulated for 5 minutes using the microscopic traffic simulation software (VISSIM). The post-simulation TPI values are calculated, and one or two schemes with a TPI improvement of 15% or more are selected as the candidate schemes to ensure theoretical validity before actual implementation.

[0032] In step S7, the V2X system conducts real-time verification and dynamic deviation correction. The roadside RSU (Rated State Unit) sends the execution plan to the connected vehicle (displaying the estimated time of passage) and the pedestrian indicator screen (displaying the waiting time for crossing) at a frequency of 2 seconds. The actual driving trajectory of the connected vehicle (uploaded every 0.5 seconds) and the pedestrian crossing behavior (detected every 5 seconds by infrared sensors) are simultaneously collected, and the deviation rate between the actual data and the predicted value is calculated (deviation rate = |actual value - predicted value| / predicted value). If the deviation rate of traffic flow in a certain direction exceeds 15% (e.g., a prediction of 100 vehicles per 5 minutes, compared to an actual value of more than 115 vehicles) or the deviation rate of pedestrian waiting time exceeds 20%, the game model is triggered to recalculate weights, dynamically adjust the green light duration (e.g., extending the main direction green light by 10 seconds) or insert a temporary phase (e.g., a 5-second pedestrian crossing transition phase) to ensure that the phase plan matches the actual traffic flow.

[0033] In step S8, a full-cycle closed-loop optimization and model iteration process uses the full 24-hour daily data as the training set, employing a transfer learning algorithm to update the game model and prediction model parameters. The source domain for transfer learning is the current intersection data, and the target domain is similar intersections within the region (with a traffic pattern match of ≥80%, e.g., similar-level main roads with similar traffic flow characteristics). Model parameters are updated weekly. The optimization effect is evaluated by the overall regional TPI improvement rate (requires a monthly average improvement of ≥5%). The updated model parameters are synchronized with the control centers of similar intersections, forming a collaborative evolutionary mechanism of "data collection - model optimization - solution sharing." This continuously improves the system's adaptability to different scenarios, ultimately achieving global optimization of intersection phase solutions within the region.

[0034] The intelligent traffic light phase optimization system based on vehicle-road collaboration runs the above-mentioned intelligent traffic light phase optimization method based on vehicle-road collaboration when the system is in use.

[0035] From the above we can know: In the present invention, through the dynamic game model, the system can perceive multi-dimensional states such as traffic flow and pedestrian waiting time in real time, dynamically adjust the demand weights of vehicles, pedestrians, and roadside facilities, and flexibly allocate green light duration - more time is given to directions with more vehicles, and priority is given to crossing needs during periods with more people, effectively shortening the average waiting time of vehicles and pedestrians, significantly improving the overall traffic efficiency of the intersection, and significantly improving the adaptability of daily traffic operations.

[0036] In the present invention, the dynamic game model quickly identifies special needs through the state-value function and dynamically increases the weight priority of emergency vehicles or pedestrians. For example, when an ambulance is detected approaching, the green light in the corresponding direction is extended first, and when pedestrians are observed waiting in a concentrated manner, a temporary crossing phase is inserted. This not only ensures efficient passage in key scenarios (such as the rapid passage of emergency vehicles), but also avoids long-term backlogs of vehicles in other directions, significantly enhancing the system's response capability and coordination level to complex and sudden scenarios.

[0037] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprises" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device that includes a series of elements includes not only those elements, but also other elements that are not explicitly listed, or also includes elements that are inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device that includes the element.

[0038] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An intelligent traffic light phase optimization method based on vehicle-road collaboration is characterized by: The method comprises the following steps: S1: Through roadside sensors, on-board equipment, and the traffic cloud platform, data such as vehicle location, speed, type, and pedestrian needs are synchronously acquired and transmitted to the regional control center in real time, providing raw materials for subsequent fusion; S2: Based on the multi-source data collected by S1, the timestamps and spatial coordinates are calibrated, and federated learning is used to fuse the local vehicle-side data and the global roadside data to generate a full-factor situation map of the intersection, providing a unified data source for status assessment; S3: Based on the situation diagram of S2, the traffic operation index is calculated using the fuzzy evaluation method from the four dimensions of flow, efficiency, fairness, and safety to identify the current phase shortcomings and provide a current basis for demand forecasting; S4: Combining the TPI from S3 with historical data, LSTM is used to predict traffic flow in all directions, pedestrian crossing times, and special vehicle needs in the next five minutes. Connected vehicles are used to obtain destination information to improve accuracy and provide input for the game model. S5: Based on the needs predicted in S4, a three-party game model of "vehicle-road-pedestrian" is constructed. With the goal of minimizing overall delay, reinforcement learning is used to dynamically adjust the weights of each party and output the phase adjustment tendency to provide a decision-making basis for solution generation. S6: Phase plan generation and screening: Based on the adjustment tendency of S5, call the preset phase template and combine it with the predicted data of S4 to generate 3 to 5 candidate plans. Use TPI pre-simulation to select the best 1 to 2 as the plans to be implemented; S7: Collaborative verification and dynamic deviation correction: Send the S6 execution plan to the connected vehicle and pedestrian screens to collect actual driving and street crossing feedback data; if the deviation exceeds 15%, trigger the S5 model to readjust the weight or insert a temporary phase; S8: Full-cycle model iteration; record the execution effect of S7, use transfer learning to update the parameters of the S5 game model and S4 prediction model, and synchronize them to similar intersections in the area to form a "collection-optimization-sharing" collaborative evolution machine.

2. The intelligent traffic light phase optimization method based on vehicle-road collaboration according to claim 1, characterized in that: In step S1, multi-source data is collected in real time through the collaboration of roadside sensing equipment, on-board terminals, and the traffic cloud platform. Roadside cameras collect image data at a frequency of 25 frames per second, covering a 500-meter range at each entrance to the intersection. Millimeter-wave radar obtains vehicle position and speed information at a scanning frequency of 0.1 seconds per time, with a ranging accuracy of ±0.1 meters. On-board OBU equipment uploads vehicle dynamic data every 2 seconds through V2X communication. Pedestrian crossing needs are detected in real time by roadside infrared sensors, and the number of waiting people and detention time are updated every 5 seconds.

3. The intelligent traffic light phase optimization method based on vehicle-road collaboration according to claim 1, characterized in that: In step S2, the spatiotemporal alignment and fusion of multi-source data are based on timestamp calibration and spatial coordinate unification; time calibration uses the NTP protocol to synchronize the clocks of all devices, and the error is controlled within ±50 milliseconds; the spatial coordinates use the center point of the intersection as the origin to establish a local coordinate system, and the relative coordinates of the roadside sensors and the GPS absolute coordinates of the on-board equipment are converted into a unified coordinate system; the fusion process adopts a federated learning framework, the vehicle-side equipment locally extracts features of the perceived local traffic data, and the roadside center aggregates all vehicle-side features and roadside global data, and updates the global model parameters every 5 minutes. The number of local iterations is set to 3 times, and finally a full-factor traffic situation map covering a 500-meter range of the intersection with an update frequency of 1 second / time is generated.

4. The intelligent traffic light phase optimization method based on vehicle-road collaboration according to claim 1, characterized in that: In step S3, the multi-dimensional evaluation of real-time traffic status is based on the integrated full-factor situation map, and quantitative analysis is carried out from four dimensions: flow, efficiency, fairness, and safety. In the flow dimension, a 5-minute sliding window is used to count the number of vehicles on each import channel to calculate the real-time flow; The efficiency dimension calculates the average delay time by comparing the actual vehicle passing time with the free flow passing time.

5. The intelligent traffic light phase optimization method based on vehicle-road collaboration according to claim 1, characterized in that: In step S4, the dynamic prediction of multi-scenario traffic demand uses an LSTM neural network as the core model, with the input being the fused data of the past 15 minutes and 64 nodes set in the hidden layer to predict the traffic status in the next 5 minutes; the model training data covers four scenarios: morning rush hour, evening rush hour, off-peak and special events on weekdays, with a training cycle of 30 days, and the prediction error is controlled within the range of ±10% of vehicle flow and ±5 seconds of pedestrian waiting time.

6. The intelligent traffic light phase optimization method based on vehicle-road collaboration according to claim 1, characterized in that: In step S5, a data-driven intelligent algorithm is used to find the optimal balance point among the conflicting demands of the three types of traffic participants: vehicle, road, and pedestrian. First, based on the traffic demand predicted in step S4 for the next five minutes, the types of entities that currently need to be coordinated are identified. Then, with the overall goal of minimizing overall delays, connected vehicles, roadside facilities, and pedestrians are defined as the three game entities, each corresponding to a set of dynamically changing strategies. Finally, a reinforcement learning algorithm is used to dynamically adjust the weights of the three parties in the overall goal, ultimately outputting the most reasonable phase adjustment tendency at the current time. The policy gradient algorithm in deep reinforcement learning is used. The state space of the model is composed of the full-factor situation map generated by S2 and the 5-minute demand data predicted by S4. The action space corresponds to the executable phase adjustment operations. The reward function is directly related to the overall goal - by calculating the overall delay reduction after executing a certain action in real time, the model is driven to continuously optimize the strategy. The calculation formula for dynamic weight allocation is: ； Among them, w i (t) represents the weight coefficient of the i-th class subject at time t, α and β are normalization coefficients, base i is the base weight of the subject, scene i (t) is the scenario correction term; this formula, through a two-layer design of "basic weight + scenario correction," ensures stability in daily scenarios while enabling rapid response to special events. The calculation formula of the state value function is: ； Among them, s represents the current traffic status, γ is the discount factor, R i (s) is the immediate benefit of the i-th agent in state s. This function directly links the weights of the three parties with their respective benefits, and the model selects the optimal action by maximizing V(s). This design enables the model's decision-making to reflect the different needs of different agents while always focusing on the overall goal of "minimizing overall delay." 7. The intelligent traffic light phase optimization method based on vehicle-road collaboration according to claim 1, characterized in that: In step S6, the intelligent generation of phase schemes and the screening of candidate sets are based on a preset basic phase template. The templates are divided according to time period characteristics: the morning peak template has a green light of 90 seconds in the main direction and 45 seconds in the secondary direction; the flat peak template has a green light of 70 seconds in the main direction and 35 seconds in the secondary direction; the special event template has a green light of 120 seconds in the main direction and 30 seconds in the secondary direction; based on the adjustment tendency output by the game model, the green light duration of the basic template is fine-tuned by ±10 seconds to generate 3 to 5 candidate schemes.

8. The intelligent traffic light phase optimization method based on vehicle-road collaboration according to claim 1, characterized in that: In step S7, the real-time verification and dynamic correction of vehicle-road collaboration sends the plan to be executed to the connected vehicle and pedestrian prompt screen through the roadside RSU device, with a sending frequency of 2 seconds per time; the actual driving trajectory of the connected vehicle and the pedestrian crossing behavior are simultaneously collected, and the deviation rate between the actual data and the predicted value is calculated.

9. The intelligent traffic light phase optimization method based on vehicle-road collaboration according to claim 1, characterized in that: In step S8, the full-cycle closed-loop optimization and model iteration use the full 24-hour data as the training set, and adopt the transfer learning algorithm to update the game model and prediction model parameters; the source domain of the transfer learning is the current intersection data, and the target domain is similar intersections in the area, and the model parameters are updated once a week.

10. The intelligent traffic light phase optimization system based on vehicle-road collaboration is characterized by: When in use, the system runs the intelligent traffic light phase optimization method based on vehicle-road collaboration as described in any one of claims 1 to 9.

Images