Intelligent remote control method and system for traffic signal based on multi-source data fusion

Abstract

This invention relates to the field of traffic signal timing control technology, and discloses a method and system for intelligent remote control of traffic lights based on multi-source data fusion. This method collects traffic flow data at intersections in real time, and completes the spatiotemporal alignment of multi-source heterogeneous data by establishing a unified Cartesian coordinate system and an event-triggered snapshot mechanism. Based on the confidence weights of visual and radar data, combined with a safety redundancy coefficient, the basic green light duration is calculated. On this basis, a downstream backpressure suppression factor and a historical residual feedback compensation term are introduced to correct the basic green light duration. Finally, a control command containing a lifecycle timestamp is generated. The mobile intelligent traffic light verifies the timeliness of the command, executes it, and uploads the actual execution data to update the next cycle's scheme. This invention introduces a downstream road segment saturation monitoring mechanism, solving the problem that traditional traffic light control strategies ignore downstream carrying capacity, and improving the robustness of traffic control and regional coordination efficiency.

Description

Technical Field

[0001] This invention relates to the field of traffic signal timing control technology, specifically to a method and system for intelligent remote control of traffic lights based on multi-source data fusion. Background Technology

[0002] Mobile intelligent traffic lights, as a flexible and efficient temporary traffic control device, are typically used when existing (fixed) traffic lights at intersections cannot function properly, necessitating temporary traffic control. Compared to fixed permanent traffic lights, they offer advantages such as self-contained power (solar / battery), no need for underground cabling, and rapid deployment, making them an important supplementary equipment for command centers to achieve regional traffic emergency evacuation and order maintenance.

[0003] In existing technologies, the control strategies for mobile traffic lights are relatively outdated, mainly relying on simple local timing control or manual on-site operation. Specifically, there are currently two main control methods: one is to preset a fixed countdown scheme for red and green lights and execute it cyclically; the other is to rely on traffic police officers on-site to use handheld remote controls to manually switch phases based on visually observed traffic flow. This control method can maintain basic traffic order during normal periods when traffic flow is stable.

[0004] However, the existing mobile traffic light control methods have some limitations in practical applications, making it difficult to cope with complex and ever-changing dynamic traffic flows. Specifically, existing fixed timing schemes are relatively rigid and cannot adapt to the characteristics of drastic fluctuations in traffic flow under temporary road conditions. They cannot dynamically adjust the control strategy for the next cycle based on actual traffic flow changes after execution, resulting in inefficiencies such as "empty traffic lights" or "car queues." Secondly, manual remote control relies on human judgment and lacks data support, making it prone to fatigue errors. In addition, mobile traffic lights often operate as information silos, lacking the ability to coordinate with the surrounding road network. That is, they only focus on the current intersection's traffic flow, ignoring the carrying capacity of downstream road segments. Frequently, if severe congestion occurs at downstream intersections, even if the current intersection's green light is on, vehicles cannot pass due to the congestion ahead, or even get stuck in the middle of the intersection, causing a "lockdown"; or the green light time is too short, leading to an increasing backlog of vehicles, ultimately causing traffic paralysis in the entire area. Therefore, there is an urgent need for intelligent remote control methods and systems for traffic lights based on multi-source data fusion to solve the above problems. Summary of the Invention

[0005] To address the problems in related technologies, this invention provides a method for intelligent remote control of traffic lights based on multi-source data fusion, thereby overcoming the aforementioned technical problems in existing related technologies.

[0006] To solve the aforementioned technical problem, the present invention is achieved through the following technical solution:

[0007] In a first aspect, embodiments of the present invention provide a method for intelligent remote control of traffic lights based on multi-source data fusion, specifically including: real-time collection of traffic flow data at the intersection through an intersection sensor network and an internet data source; establishing a unified Cartesian coordinate system and an event-triggered snapshot mechanism for the intersection, performing spatiotemporal alignment processing to spatially map the traffic flow data to the same physical model of the intersection, and synchronizing it temporally based on the start time of the signal control cycle; and calculating the traffic flow data based on the spatiotemporally aligned processing, using visual and radar reliability weighting coefficients, combined with lane saturation dissipation speed and maximum capacity, to meet the current traffic demand at the intersection. The system calculates the base green light duration; it introduces a downstream backpressure suppression factor and a historical residual feedback compensation term to construct a duration correction model, which corrects the base green light duration to generate a corrected green light execution duration; the downstream backpressure suppression factor is used to reduce the passage time when downstream is congested, and the historical residual feedback compensation term is used to compensate for the residual queue from the previous cycle; the cloud server encapsulates the corrected green light execution duration into a control command with a lifecycle timestamp and sends it to the mobile intelligent traffic signal; the mobile intelligent traffic signal verifies the timeliness of the command and executes it, and collects residual queue data at the end of the cycle and sends it back to the cloud for use in correcting the scheme for the next cycle.

[0008] As a preferred embodiment of the intelligent remote control method for traffic lights based on multi-source data fusion described in this invention, the traffic flow data includes the real-time queue length of each lane. Vehicle dynamic arrival rate Vehicle departure speed and the real-time saturation of downstream road sections ;

[0009] The real-time queue length of each lane By acquiring key frames of surveillance video through high-definition cameras, constructing an inverse perspective transformation matrix to map the image pixel coordinates to road world coordinates, and identifying the Euclidean distance between the farthest vehicle's rear end and the stop line while waiting at a red light;

[0010] The vehicle dynamic arrival rate The system tracks the trajectory of moving targets entering the field of view using millimeter-wave radar, counts the number of vehicles whose velocity vectors point towards the intersection within a set time window, and calculates the average instantaneous speed of the target crossing the stop line during the green light phase to obtain the vehicle's departure speed. ;

[0011] The real-time saturation of the downstream road section By separately obtaining the saturation of the congestion delay index conversion of Internet map services Saturation calculated by the geomagnetic induction coil at the downstream intersection Then take the maximum of the two to obtain, that is .

[0012] As a preferred embodiment of the intelligent remote control method for traffic lights based on multi-source data fusion described in this invention, the spatiotemporal alignment processing includes: projecting the visually recognized vehicle coordinates and the radar target coordinates transformed by a rotation and translation matrix onto a unified Cartesian coordinate system using a homography matrix, and performing spatial position matching; and using the signal control cycle... The start time Open time window as a benchmark Take the arithmetic mean of the radar and visual data within the window, and correlate it with the distance. The most recent valid internet data This is used to ensure the synchronization of traffic flow data over time.

[0013] As a preferred embodiment of the intelligent remote control method for traffic lights based on multi-source data fusion described in this invention, the formula for calculating the basic green light duration is as follows:

[0014] ;

[0015] In the formula, This indicates the control cycle that is about to be executed. Based on the basic green light duration, The lane saturation dissipation velocity, To maximize the capacity of the lanes, This is the visual reliability weighting coefficient. This is the radar confidence weighting coefficient. Based on the basic safety redundancy factor, In order to match the current control cycle Historical average green light duration for items with the same date attribute and within the same statistical window.

[0016] As a preferred embodiment of the intelligent remote control method for traffic lights based on multi-source data fusion described in this invention, the lane saturation dissipation speed... The update is performed using an exponentially weighted moving average algorithm, and the update formula is as follows:

[0017] ;

[0018] In the formula, The vehicle departure speed for the current cycle. To update the coefficients, This is the minimum flow rate threshold.

[0019] As a preferred embodiment of the intelligent remote control method for traffic lights based on multi-source data fusion described in this invention, the visual reliability weight coefficient... The radar confidence weighting coefficient is the arithmetic mean of the confidence scores of the queued vehicles identified in the current period. The coefficient of determination for the linear regression of the distance and time trajectory of the radar-tracked target within the current period. The root mean square value.

[0020] As a preferred embodiment of the intelligent remote control method for traffic lights based on multi-source data fusion described in this invention, the expression of the duration correction model is:

[0021] ;

[0022] In the formula, To adjust the green light execution time, This is the downstream back pressure suppression factor term. The sensitivity coefficient, The congestion trigger threshold, For historical residual feedback compensation items, For feedback gain coefficient, This represents the residual queue length at the end of the previous cycle.

[0023] As a preferred embodiment of the intelligent remote control method for traffic lights based on multi-source data fusion described in this invention, the verification of the timeliness of the instruction includes: a cloud server generating a timestamp containing the instruction's lifecycle. and effective window The mobile intelligent traffic light receives the control message and verifies the current time. The difference: If the difference is less than Then the calculated green light execution time will be executed. If the difference is greater than or equal to If the command is not executed, the instruction will be discarded and the preset minimum green light time will be executed.

[0024] As a preferred embodiment of the intelligent remote control method for traffic lights based on multi-source data fusion described in this invention, the scheme for correcting the next cycle includes: at the end of the current cycle, the mobile intelligent traffic light calls a visual sensor to calculate the remaining queue length. Record the actual green light duration. It is then uploaded to the cloud to update the historical residual feedback compensation items of the next cycle duration correction model.

[0025] Secondly, embodiments of the present invention provide an intelligent remote control system for traffic lights based on multi-source data fusion, comprising: a multi-source perception module, used to collect real-time queue length, vehicle dynamic arrival rate, vehicle departure speed, and real-time saturation of downstream road sections for each lane through an intersection sensor network and Internet data sources; a spatiotemporal alignment module, used to establish a unified coordinate system and an event-triggered snapshot mechanism for the intersection, to perform spatial location mapping and time reference synchronization of the collected heterogeneous traffic flow data, thereby achieving multi-source data fusion; an intelligent calculation module, used to calculate the basic green light duration based on the fused data, and to construct a duration correction model by combining the downstream back pressure suppression factor and the historical residual feedback compensation term, dynamically generating the final green light execution duration; and a remote control module, used to generate and issue control commands containing lifecycle timestamps, drive the traffic lights to execute after verifying the timeliness of the commands, and collect residual queue data after execution for closed-loop feedback.

[0026] The present invention has the following beneficial effects:

[0027] 1. This invention introduces a downstream road segment saturation monitoring mechanism. When the system detects that the downstream road segment is becoming congested, it uses nonlinear suppression logic to automatically reduce the green light duration at the current intersection. By stopping the passage, it reduces the number of vehicles heading to the downstream road segment, thus avoiding more serious congestion. At the same time, it can prevent vehicles from getting stuck in the middle of the intersection due to congestion in the downstream road segment, which helps to ensure smooth traffic in other directions at the current intersection.

[0028] 2. The mobile intelligent traffic light of this invention establishes a closed-loop control mode of remote calculation, local execution, and feedback correction. This allows the system to automatically assess the residual queue length of the previous cycle at the end of each cycle and convert it into a time compensation amount, which is then added to the green light duration of the next cycle. In other words, if the residual queue length of the previous cycle is long, the green light duration of the next cycle is increased accordingly. This feedback compensation mechanism based on historical residual queue length enables the system to adaptively eliminate control errors, helping to prevent a continuous increase in traffic congestion caused by insufficient green light time.

[0029] 3. Considering the network jitter or delay risks that remote control may face, this invention incorporates a lifecycle timestamp and valid window verification mechanism into the control commands. This allows the mobile intelligent traffic light to perform real-time verification when receiving commands. Once a command is found to have expired due to transmission delay, it will be automatically discarded and switched to the preset minimum green light time. This ensures that even under poor communication quality, intersection signal control can still guarantee the basic traffic safety of vehicles and pedestrians, avoiding traffic accidents caused by the execution of delayed commands.

[0030] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0032] Figure 1 The present invention provides a flowchart of a method for intelligent remote control of traffic lights based on multi-source data fusion.

[0033] Figure 2 This is a schematic diagram of the S5 process provided by the present invention.

[0034] Figure 3 The overall system architecture diagram provided by this invention.

[0035] Figure 4 This is a schematic diagram of a downstream back pressure suppression scenario provided by the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Example 1

[0038] Existing mobile intelligent traffic lights either rely on manual remote control operation or are set to execute a fixed countdown without any intervention. This approach is particularly inflexible in practical applications. Common problems include: if severe traffic congestion occurs at the downstream exit of the temporarily controlled section, even if the mobile intelligent traffic light is currently on green, vehicles cannot pass due to the congestion ahead and are stuck in the middle of the temporary intersection, causing a "lockdown"; or the green light time is too short, leading to an ever-increasing backlog of vehicles and ultimately causing traffic paralysis on that section of the road.

[0039] To solve the above technical problems, such as Figure 1As shown, Embodiment 1 of the present invention provides a method for intelligent remote control of traffic lights based on multi-source data fusion. Specifically, Embodiment 1 takes an intersection of a main road and a secondary road requiring temporary traffic control as an example: A mobile intelligent traffic light, a 4-megapixel AI capture camera (model HV-400, coverage range 150 meters), and a 77GHz millimeter-wave radar (model MW-77, supporting all-weather trajectory tracking) are deployed on-site. When the existing traffic lights at the intersection require maintenance, the traffic command center needs to implement intelligent remote control of the traffic lights during the morning rush hour using the mobile intelligent traffic light to cope with sudden traffic congestion. The mobile intelligent traffic light can be controlled by a control handle, integrates a windproof and stable chassis and a lifting mast, is equipped with a dual-mode independent power supply system of solar energy and batteries, and has a built-in high-precision timing module, edge computing module, and 5G communication module.

[0040] In the specific implementation of Example 1: First, traffic flow data at the intersection is collected in real time through an intersection sensor network and an internet data source. This method overcomes the blind spots and information silos of a single sensor by integrating visual perception of existing queuing, radar perception of incremental arrivals, and internet perception of downstream exit constraints, thereby quantifying the traffic demand at the intersection. Second, a unified Cartesian coordinate system and an event-triggered snapshot mechanism are established for the intersection, and spatiotemporal alignment processing is performed to spatially map the traffic flow data to the same physical model of the intersection and temporally synchronize it based on the start time of the signal control cycle. This method eliminates the spatial parallax caused by different installation locations of heterogeneous sensors and the time lag caused by differences in data update frequency, ensuring that all input variables belong to the same spatiotemporal slice in a physical sense, providing a homogeneous data foundation for accurate timing. Then, based on the traffic flow data processed by spatiotemporal alignment, the basic green light duration that meets the current intersection traffic demand is calculated using visual and radar reliability weight coefficients, combined with lane saturation dissipation speed and maximum capacity. This method introduces a dynamic weighted reliability evaluation mechanism and adaptive flow velocity parameters, which can automatically adjust sensor dependence according to environmental interference (such as rain and fog), so that the calculated green light duration not only conforms to the current physical traffic limits but also has extremely high anti-interference robustness. Subsequently, a downstream backpressure suppression factor and a historical residual feedback compensation term are introduced to construct a duration correction model to correct the basic green light duration and generate a corrected green light execution duration. This invention prevents vehicles from blindly entering the intersection during downstream congestion, causing a "blockage" and affecting the passage of vehicles in other directions, by using nonlinear suppression logic, and eliminates accumulated errors by using a residual feedback mechanism, realizing the transformation of the control strategy from single-point local optimization to regional collaborative safety. Finally, the cloud server encapsulates the corrected green light execution duration into a control command with a lifecycle timestamp and sends it to the mobile intelligent traffic signal light. The mobile intelligent traffic signal light executes the command after verifying its timeliness, and collects residual queuing data at the end of the cycle and sends it back to the cloud for use in correcting the scheme for the next cycle. This invention uses a command lifecycle verification mechanism to avoid the risk of outdated operations caused by network jitter, ensuring guidance safety under communication failures, and endows the system with the ability to continuously self-optimize through a complete closed loop of perception, calculation, execution, and feedback.

[0041] Furthermore, to better illustrate the technical solution of Embodiment 1 of the present invention, a detailed description of the intelligent remote control method for traffic lights based on multi-source data fusion is provided, specifically including the following:

[0042] S1. The system collects traffic flow data at the intersection in real time through a multi-source heterogeneous sensor network. The traffic flow data includes: real-time queue lengths for each lane. Dynamic arrival rate of vehicles and vehicle departure speed Real-time saturation of downstream road sections . Specifically:

[0043] The real-time queue length of each lane The following sub-steps are performed using high-definition cameras deployed at intersections to acquire data: First, keyframes of the surveillance video are acquired; then, a lightweight convolutional neural network is used to identify vehicle targets in the image, outputting the bounding box coordinates of the vehicles; finally, based on the camera's installation height... With pitch angle Construct an inverse perspective transformation matrix to transform the pixel coordinates of the two-dimensional image. Mapped to physical coordinates in the road surface world coordinate system This eliminates near-to-far distortion in the image; finally, in the transformed top view, virtual detection areas (ROIs) are set along the center lines of each lane, and the system identifies the rear coordinates of the furthest vehicle waiting at a red light. Calculate the coordinates of this coordinate and the coordinates of the parking line. The Euclidean distance between them is the current queue length. It should be noted that if a queue overflows the field of view, then... Marked as maximum visible distance.

[0044] The dynamic arrival rate of the vehicle and vehicle departure speed The following measurements were performed using the frequency-modulated continuous wave (FMCW) characteristics of a 77 GHz millimeter-wave radar: First, the radar calculated the target's range and radial velocity by transmitting a linear frequency-modulated wave and receiving the echo; then, using the extended Kalman filter (EKF) algorithm, a unique TrackID was assigned to each moving target entering the radar's field of view (FOV, typically 150-200 meters), and its trajectory was continuously tracked; then, at the distance to the parking line... A virtual counting section is set at a distance of 100 meters (for example), and the system counts the data within a set time window. The number of unique TrackIDs that passed through the cross-section and whose velocity vectors pointed towards the intersection within the past 10 seconds (e.g., the number of times they would arrive at the stop line) is calculated using linear regression, and the total number of vehicles expected to arrive is obtained by summing these data. For the green light phase, the system locks onto the trajectory of the target vehicle that has crossed the stop line and entered the intersection area, calculates the arithmetic mean of its instantaneous speed, and uses this as the average vehicle departure speed. It is used to correct the saturation dissipation rate parameter.

[0045] The real-time saturation of the downstream road section The strategy of primarily using internet data and supplementing it with geomagnetic correction is employed, and the method is as follows: First, the system... Every few seconds (e.g., 30 seconds), the system calls the traffic status APIs of internet map service providers such as Gaode and Baidu to obtain the Congestion Delay Index (TTI) of downstream road segments, and then uses a normalized formula to determine the saturation of the internet map. This is converted into a preliminary saturation value in the 0-1 range; when there is a delay or missing data in the internet data, data from the geomagnetic induction coil installed at the stop line of the downstream intersection entrance is read, the time occupancy rate is calculated, and the geomagnetic saturation is determined. ,in This represents the total time the detector is occupied by vehicles. For the statistical period; the final real-time saturation of the downstream road section. Take the maximum of the two, that is This ensures highly sensitive detection of congestion risks and prevents misjudgments caused by distortion from a single data source.

[0046] In this embodiment 1, the real-time queue length of each lane is obtained through a high-definition camera. This allows for the quantification of current traffic demand at intersections; millimeter-wave radar is used to capture the dynamic arrival rate of vehicles. This allows for the quantification of incremental demand at current intersections, preventing a large influx of vehicles from arriving immediately after the green light turns off; simultaneously, it utilizes internet data to obtain real-time saturation levels of downstream road sections. This allows for the quantification of the current exit constraints at the intersection.

[0047] S2. The traffic flow data above undergoes spatiotemporal alignment processing, specifically including the following sub-steps:

[0048] S21. Due to the different spatial references of the data from cameras, radar, and electronic maps (image pixel coordinates, polar coordinates, and GPS coordinates, respectively), the system first establishes a unified Cartesian coordinate system for the intersection, with the midpoint of the stop line as the origin and the lane travel direction as the Y-axis. Specifically, using a calibrated homography matrix, the system uses the bottom center point of the vehicle bounding box captured by the camera... Projected to a unified coordinate system And according to coordinates The value determines the lane ID of the vehicle. ); the polar coordinates of the target output by the millimeter-wave radar The data is converted to Cartesian coordinates, aligned to the same coordinate system using a coordinate rotation and translation matrix, and then matched with the spatial location of the visually recognized target. If a match is found, the attributes of both are merged (visual recognition determines the category, and radar determines the speed). The road-level congestion index obtained from the internet map API is then associated with the corresponding entrance lane group through a topological mapping table. For example, "Congestion on Jiefang Road from west to east" is mapped to the "straight lane" and "right turn lane" at this intersection.

[0049] S22. To address the issue of inconsistent data update frequencies from different sensors, the system employs an event-triggered snapshot mechanism for synchronization: the cycle is controlled by a signal. The start time Based on this, open a time window (e.g., 200ms); take the arithmetic mean or latest value of radar and visual data within this window as the input for the current period; for the Internet with a low update frequency Data, using distance The most recent API response ensures the causal validity of the data over time.

[0050] In this embodiment 1, deep fusion of multi-source data is achieved by establishing a unified Cartesian coordinate system and an event-triggered snapshot mechanism. Specifically, for example: assuming the current signal control cycle... The start time for Preset time snapshot window for In this window ( to ), millimeter-wave radar with Four sets of reports about the target vehicle were submitted at intervals. Speed ​​data While visual sensors are only A frame of location data was uploaded at a specific time. The system performed the following operations: First, the radar's polar coordinate data was mapped to a Cartesian coordinate system using a rotation and translation matrix, and it was found that its Euclidean distance from the visual recognition coordinates was only [missing information]. (If the value is less than the matching threshold), it is confirmed that the two are the same target; secondly, the arithmetic mean of the radar velocities within the window is calculated. As the effective speed input for this cycle Finally, for slowly updating internet data, the system automatically retrieves... The most recent valid record before the time, for example Obtained downstream saturation Then the current moment will continue to be maintained. This method achieves spatiotemporal alignment of all input variables. It spatially maps high-frequency perception data from vision and radar to the same intersection physical model and temporally aligns them using the start time of the signal control cycle as the anchor point. Simultaneously, it associates low-frequency internet traffic data with the current cycle using the temporal nearest neighbor principle. This spatiotemporal synchronization eliminates parallax caused by different sensor installation locations and time lags caused by differences in data update frequencies, ensuring that all parameters input to the timing calculation model belong to the same spatiotemporal slice in a physical sense. This provides accurate and isomorphic data support for subsequent calculations of the basic green light duration.

[0051] S3. Calculate the basic green light duration to meet the current intersection traffic demand based on traffic flow data processed by spatiotemporal alignment. The calculation formula is as follows:

[0052] ;

[0053] in, Indicates the control cycle that is about to be executed; The queue length as perceived visually; This refers to the lane saturation dissipation speed; The estimated number of vehicles arriving, as sensed by radar; This represents the maximum traffic capacity of the lane, measured in veh / s, where veh refers to the number of vehicles and s refers to seconds. and The preset reliability weight coefficients, Based on the basic safety redundancy factor, In order to match the current control cycle Historical average green light duration for items with the same date attribute and within the same statistical window.

[0054] Specifically, considering that weather (such as rain and snow), lighting conditions, or driver behavior can affect vehicle traffic efficiency, the system uses an exponentially weighted moving average (EWMA) algorithm to calculate lane saturation dissipation speed. Updates can be performed using the following formula:

[0055] ;

[0056] Simultaneously set a lower limit constraint:

[0057] ;

[0058] In the formula, These are the saturated dissipation rate parameters used for calculations in this period; The vehicle departure speed determined in step S1; To update the coefficients; This is the minimum flow rate threshold.

[0059] For example:

[0060] In this embodiment 1, the update coefficients are... To avoid severe system oscillations caused by fluctuations in a single measurement (such as a novice driver starting slowly), the method for determining this is as follows: The system calculates the coefficient of variation of all vehicle speed samples collected by the radar within the current period. Using inverse proportional functions Determine and update the weights. When traffic flow is stable, i.e. When it approaches 0, When the value approaches 1, the system responds quickly to current changes; when traffic flow becomes chaotic, that is... When it increases, By reducing the data quality, the system tends to maintain the stability of the historical model, thus achieving adaptive filtering based on data quality.

[0061] In this embodiment 1, the minimum flow velocity threshold This characterizes the physical critical speed at which the lane can maintain continuous saturation traffic, preventing computational overflow caused by an excessively small denominator due to congestion. Its determination method is as follows: The system retrieves historical flow velocity data from the past 30 natural days from the database and constructs a normal distribution model. According to statistics The guidelines will Determined as If the calculated result is lower than the physical limit, the minimum continuous flow velocity derived based on industry standards (such as the Road Capacity Manual) shall be used. As a safety net, thus ensuring the physical objectivity of the lower limit, among which For standard vehicle length, (This is the maximum saturation time interval).

[0062] In this embodiment 1, the visual reliability weighting coefficient The determination method is as follows: The system extracts the confidence score of each vehicle identified by the convolutional neural network in the current period. Calculate the confidence score The arithmetic mean of the values ​​is used as the visual reliability weighting coefficient, i.e. , This represents the total number of vehicles in the queue identified by the convolutional neural network within the current period; the visual reliability weight coefficient. This characterizes the system's level of confidence in the visual perception results: when lighting conditions are good and occlusion is minimal, the average confidence level is [value missing]. High, The confidence level approaches 1, indicating that the system fully trusts visually measured queue lengths; however, when encountering heavy rain or strong backlighting that causes a decrease in confidence, The system automatically reduces the weight of visual items to suppress potential false detection risks and prevents green light time from being assigned to non-existent "phantom vehicles".

[0063] In this embodiment 1, the radar confidence weighting coefficient The determination method is as follows: The system performs linear regression analysis on the continuous spatiotemporal trajectory points of each tracked target within the radar field of view, and extracts its coefficient of determination. The root mean square value of the entire sample is calculated as the radar confidence weighting coefficient, i.e. , The total number of vehicles expected to arrive as counted by the radar in the current period (i.e. ), Indicates the first The goodness of fit between the distance and time trajectory of the target vehicle is determined by the analysis of the first target vehicle. One goal in the past Inside the window Linear regression was performed on each location point, and the square of the Pearson correlation coefficient was calculated.

[0064] It should be noted that the visual reliability weighting coefficient and radar confidence weighting coefficient The update frequency is strictly synchronized with the signal control cycle frequency. Specifically, in each control cycle... Basic green light duration Before the calculation instruction is initiated (i.e. At a time of microseconds, the system triggers a weight update operation.

[0065] In this embodiment 1, the historical average green light duration The determination method is as follows: The system first retrieves all historical samples from the cloud database that have the same date attribute as the current control period and are within the same standard 15-minute statistical window, based on the principle of time isomorphism, and retains only the visual confidence score. High-quality sensory records approaching 1 were obtained; subsequently, historical queuing and clearing times were calculated for the initial screening samples, and the interquartile range (IQR) rule was applied for objective cleaning, by calculating the lower quartile of the sample distribution. and the upper quartile Automatically remove items falling within the range Statistical outliers other than those identified during the cleaning process are then identified to obtain the effective sample set. Finally, the arithmetic mean method is used to calculate the central tendency of the remaining effective sample set after cleaning, thereby obtaining the... value.

[0066] In this embodiment 1, the basic safety redundancy coefficient The method for determining the accuracy of the system, which is used to eliminate the impact of environmental changes (such as slower vehicle start-up in rain or snow), is as follows:

[0067] First, in the During each control cycle, the millimeter-wave radar deployed on the edge locks onto the furthest vehicle in the queue, i.e., the last vehicle in the queue. When this last vehicle crosses the stop line, the system records the time difference from when the green light turns on to that moment, which is recorded as the actual clearing time. At the same time, the system reads the theoretically calculated duration of that cycle. .

[0068] Then, calculate the clearance ratio for the current cycle. :

[0069] ;

[0070] like This indicates that the actual traffic efficiency is lower than the theoretical value (e.g., due to slippery road surfaces or slow driver reaction), requiring additional redundancy; if This demonstrates that the theoretical model is highly accurate.

[0071] Finally, the system maintains a capacity of (For example A first-in-first-out queue (corresponding to the past 30 cycles, this value is objectively determined by the memory space) will be used. The latest... Store in the queue. Calculate the queue. Arithmetic mean of the samples and sample standard deviation Then the coefficient of the next period based on The criterion can be determined as a confidence upper bound that covers 99.73% of the common occurrences, namely: .

[0072] Specifically, for example: setting , , The historical average green light duration for that period is set based on historical data statistics. 30 seconds. Assume the visual sensor detects the queue length of the straight-ahead lane. The current saturation dissipation rate corrected by the EWMA algorithm Millimeter-wave radar predicts the number of vehicles expected to arrive. Vehicles, maximum lane capacity veh / s (corresponding to a saturation flow rate of 1800veh / h). The system then calculates the basic green light duration. This result indicates that the basic green light requirement for this intersection in this cycle is... This duration ensures that the current 80-meter queue of vehicles is cleared, while also allowing sufficient time to accommodate the upcoming arrival of 9 vehicles.

[0073] S4. To eliminate the cumulative error that may be generated by open-loop control and to prevent blindly releasing traffic when downstream is congested, a historical residual feedback compensation term and a downstream back pressure suppression factor term are introduced to correct the base duration, generating a corrected green light execution duration. Specifically, it includes the following steps:

[0074] S41, The system reads the previous cycle ( Residual queue length at the end This is considered as the system's steady-state error; and the residual queue length is converted into the required compensation time. Then, a feedback gain coefficient is introduced. The historical residual feedback compensation item was determined as follows: ;

[0075] Furthermore, based on the basic green light duration Calculate the corrected unconstrained duration: .

[0076] S42. When determining the green light duration at an intersection, not only the current traffic capacity of the intersection but also the traffic capacity of downstream road segments must be considered. If congestion occurs in downstream road segments, i.e. If the current value approaches 1, forcibly allowing traffic to pass through the intersection would cause vehicles to overflow into the center of the intersection, leading to severe congestion. Therefore, a downstream backpressure suppression factor is introduced to construct a nonlinear suppression function. Adjust duration.

[0077] In this embodiment, the nonlinear suppression function A variant of the Logistic function is chosen, utilizing its smooth step characteristic near the threshold to achieve a soft handover from full-speed clearance to cut-off clearance. The specific nonlinear suppression function is constructed as follows:

[0078] ;

[0079] In the formula, Sensitivity coefficient; This is the congestion trigger threshold. The formula represents: when... When downstream flow is unimpeded, the exponential term approaches 0. , indicating no suppression; when That is, when downstream congestion occurs, the exponential term approaches infinity. A value approaching 0 indicates a forced cutoff;

[0080] For example:

[0081] In this embodiment 1, the sensitivity coefficient The determination method is as follows: set the downstream saturation to reach the physical limit. At that time, inhibitory factor Forced decay to the cutoff threshold is required. Boundary conditions Substituting the suppression function, we can calculate... .in A value of 0.98 is typically used, corresponding to a road segment that is almost entirely occupied; cutoff threshold A value of 0.01 indicates that pedestrians are allowed to cross the road for only 1% of the extremely short green light period.

[0082] In this embodiment 1, the congestion trigger threshold The critical physical point characterizing the phase transition from free flow to congested flow in the downstream road segment is determined as follows: Based on historical monitoring data of the downstream road segment, the system constructs a basic flow-saturation (QS) graph model, and solves the equation... The extreme points are used to determine the critical saturation level at which the road segment can maintain maximum traffic capacity. and set it as .

[0083] S43. The downstream back pressure suppression factor term obtained in step S42 is... As a multiplicative coefficient, it is applied to the unconstrained duration obtained in step S41 to construct a duration correction model and generate the corrected green light execution duration. :

[0084] .

[0085] In this embodiment 1, the synergistic effect of nonlinear suppression and historical feedback compensation mechanisms achieves a shift from a single-point maximum passage strategy to a regional collaborative control strategy. (Refer to...) Figure 4 The diagram illustrates a downstream backpressure suppression scenario. The system utilizes the smooth step characteristic of the Logistic function to intervene early when the saturation of the downstream road segment approaches a threshold. By reducing the green light time at the current intersection, it prevents more vehicles from driving downstream and causing more severe congestion; simultaneously, it avoids a "deadlock" at the current intersection, which would affect the normal passage of vehicles in other directions. Specifically, for example: assuming the system has a preset sensitivity coefficient... Congestion trigger threshold Feedback gain coefficient The basic green light demand calculated at the current moment. Multi-source sensing data shows that the real-time saturation of the downstream road section is Although it did not exceed the 0.85 threshold, it was already operating at a high level, and there was residual queue length in this lane from the previous cycle. (Approximately 3 vehicles), saturation dissipation rate Substitute the values ​​into the duration correction model to calculate the corrected green light execution time. Round down to the nearest integer: 26 This result indicates that, given the potential congestion risk faced by downstream road sections ( The system did not execute in full. Instead of using a base time of 2 seconds, the time correction model pre-suppresses approximately 22% of the traffic flow, resulting in a final 26-second solution that achieves an optimal balance between preventing downstream oversaturation and ensuring basic traffic flow at the current intersection.

[0086] S5, Remote Command Issuance and Closed-Loop Feedback Execution (Refer to...) Figure 2 As shown, the specific steps include:

[0087] S51, The cloud server will calculate the result from step S3. The data is encapsulated into control messages conforming to GB / T20999-2017 "Data Communication Protocol between Traffic Signal Controller and Host Computer". Furthermore, to address potential network jitter in remote communication, the system embeds a command lifecycle timestamp into the message. The specific message includes... ,in, This indicates a globally unique instruction code generated by the cloud control server. This indicates the revised green light duration. The specific signal light group number it operates on, This represents the absolute system time at which the cloud server encapsulates and issues the instruction. This indicates the maximum allowable delay threshold preset by the cloud algorithm based on the urgency of the current control task.

[0088] When the mobile intelligent traffic light receives a command, it performs a time delay check: first, it checks the current time against... Is the difference less than If the value is less than 20 seconds, the delay check "passes"; otherwise, the delay check "fails", indicating that the instruction has expired. The signal controller will automatically discard the remote instruction and force the execution of the preset minimum green light time (e.g., 20 seconds) to ensure that vehicles and pedestrians at the front of the stop line can start and cross the street safely, while preventing the execution of delayed timing errors due to network latency.

[0089] S52. When the delay check "passes", If the time is less than the preset minimum green light time, the preset minimum green light time will be enforced; otherwise, the default green light time will be executed. .

[0090] Furthermore, to avoid driver confusion caused by sudden changes in the countdown timer, the traffic signal employs a smooth transition strategy. :

[0091] like If the countdown is greater than the current countdown, the traffic signal will seamlessly transition when there are 5 seconds left in the countdown by extending the last second of the display.

[0092] At the start of command execution, the signal device sends an "execution confirmation" signal to the cloud via a heartbeat packet, marking the official start of control for this cycle.

[0093] S53, at the moment the green light ends in the current cycle, i.e. End of period The system immediately triggers a post-evaluation mechanism: it uses the visual sensor to capture images of residual vehicles behind the current parking line and calculates... Record the actual duration of the green light. ;Will Package and upload to the cloud database.

[0094] After receiving this feedback, the cloud algorithm module uses it as... The input parameters of the cycle, the historical residual feedback compensation terms of the update duration correction model, complete the complete closed loop from remote calculation to local execution and back to remote correction.

[0095] In this embodiment 1, by embedding a lifecycle timestamp into the control command and performing valid window verification, the risk of outdated command execution caused by network jitter or delay is avoided by using timestamp verification, ensuring fault-oriented safety when communication is unstable; at the same time, by real-time transmission of residual queuing data after execution to correct the timing scheme of the next cycle, the system can adaptively eliminate model prediction errors and execution deviations, thereby ensuring the reliability and accuracy of remote traffic signal control in complex network environments and dynamic traffic flows.

[0096] Example 2

[0097] As a second embodiment of the present invention, refer to Figure 3 The system architecture diagram shown in this embodiment, based on embodiment 1, also discloses an intelligent remote control system for traffic lights based on multi-source data fusion, specifically including: a multi-source sensing module, a spatiotemporal alignment module, an intelligent computing module, and a remote control module.

[0098] The multi-source sensing module is used to collect real-time queue lengths for each lane through intersection sensor networks and internet data sources. Vehicle dynamic arrival rate Vehicle departure speed and the real-time saturation of downstream road sections Specifically, this module integrates a lightweight convolutional neural network built into the edge computing module for identifying vehicle targets in images.

[0099] The spatiotemporal alignment module is used to establish a unified coordinate system for intersections and an event-triggered snapshot mechanism, and to map the collected heterogeneous traffic flow data to spatial locations and synchronize them with a time reference, thereby achieving multi-source data fusion.

[0100] The intelligent computing module is used to calculate the basic green light duration based on the fused traffic flow data. And combined with downstream back pressure inhibition factor terms A duration correction model is constructed by combining historical residual feedback compensation items to dynamically generate the final green light execution duration. ;

[0101] The remote control module is used to generate and issue control commands containing lifecycle timestamps, drive the traffic lights to execute after verifying the timeliness of the commands, and collect residual queuing data after execution for closed-loop feedback.

[0102] In the specific implementation of Implementation 2 above, firstly, the multi-source perception module obtains the real-time queue length of each lane through high-definition cameras. This allows for the quantification of current traffic demand at intersections; millimeter-wave radar is used to capture the dynamic arrival rate of vehicles. This allows for the quantification of incremental demand at current intersections, preventing a large influx of vehicles from arriving immediately after the green light turns off; simultaneously, it utilizes internet data to obtain real-time saturation levels of downstream road sections. This quantifies the exit constraints at the current intersection. Secondly, the spatiotemporal alignment module establishes a unified Cartesian coordinate system with the midpoint of the intersection's stop line as the origin. It projects visual pixel coordinates to physical coordinates using a homography matrix and aligns radar targets to the same coordinate system using a rotation and translation matrix. In the time dimension, a time window is opened with the start of the signal control cycle as the anchor point, and an event-triggered snapshot mechanism is used to synchronously sample multi-source data. This method eliminates sensor parallax and transmission lag, ensuring that variables input to the algorithm model belong to the same spatiotemporal slice, thus solving the problem of spatial and temporal asynchrony in heterogeneous traffic flow data. Then, the intelligent computing module not only considers the queue length at the current intersection... and arrival rate The calculation of basic requirements also incorporates the real-time saturation of downstream road sections. As a constraint, when downstream congestion is detected, the Logistic nonlinear suppression function is used. Automatically reduce the green light duration at the current intersection and incorporate residual queues from the previous cycle. Time compensation is implemented. This dynamic calculation mechanism effectively prevents vehicles from "clogging up" at intersections and can adaptively eliminate system errors, realizing a shift in control strategy from local optima to regional collaboration. Finally, the remote control module addresses network jitter issues in remote control by embedding lifecycle timestamps and effective windows into control commands. The mobile intelligent traffic lights perform real-time verification when receiving commands; if a command expires due to transmission delay, it is automatically discarded and switched to a preset safe timing scheme. Simultaneously, the system triggers a post-evaluation mechanism at the end of the cycle, transmitting the actual execution results back to the cloud, forming a complete closed loop of perception, calculation, execution, and feedback, ensuring continuous optimization and security of intelligent control.

[0103] Furthermore, a lightweight convolutional neural network is provided for recognizing vehicle targets in images and outputting the bounding box coordinates of the vehicles. In this invention, the lightweight convolutional neural network uses an improved YOLOv8-Nano and mainly consists of three parts: a backbone feature extraction network, a neck feature fusion network, and a decoupled head.

[0104] Backbone: A lightweight variant of CSP-Darknet (Cross-Stage Local Network) is used as the backbone. To adapt to edge computing power, standard convolutional layers are replaced with depthwise separable convolutions, i.e., first... Channel-wise convolution extracts spatial features, and then through Channel fusion is performed using pointwise convolutions. This structure reduces the computational cost to one-third that of standard convolutions while maintaining feature extraction capabilities. Left and right. The backbone network outputs feature maps at three scales at different depths. , respectively corresponding (Detecting small targets, such as vehicles queuing in the distance) (Target under detection) and (Resolution for detecting large targets, such as nearby vehicles).

[0105] The Neck Feature Fusion Network employs a PANet (Path Aggregation Network) structure. It includes top-down upsampling paths and bottom-up downsampling paths. The top-down paths transmit high-semantic features from deeper layers to shallower layers, enhancing vehicle category recognition; the bottom-up paths transmit high-localization-accuracy features from shallower layers to deeper layers, enhancing the regression accuracy of vehicle bounding boxes. This bidirectional fusion mechanism ensures robust detection of vehicles at different distances and of different sizes at intersections.

[0106] Decoupled Detection Head: Unlike traditional coupled heads, this lightweight convolutional neural network employs a decoupled structure, splitting the "classification task" and "regression task" into two independent convolutional branches at the end of the feature map.

[0107] Regression branch: Focuses on predicting the coordinate offset of the vehicle bounding box, outputting... dimensional vector This is directly used in step S1 to calculate the real-time queue length of each lane. .

[0108] Classification confidence branch: Focuses on predicting the probability that the target belongs to "vehicle", the output is between arrive confidence scores between This decoupling design resolves the conflict between classification and regression tasks in the feature space, resulting in a higher confidence score in the output. This more purely reflects the model's degree of confirmation of the target, thus directly supporting the use of [the model's capabilities] in step S3. Calculate the visual reliability weight coefficient The logical rationality of the vehicle's contribution to the overall confidence level is as follows: the higher the confidence level, the larger the denominator of the weights, and the more robust the vehicle's contribution to the overall confidence level.

[0109] Output data format: The lightweight convolutional neural network ultimately outputs a set of tensors. Each row represents a detected vehicle target, containing: ,in, The coordinates of the vehicle's center point. For the width and height of the vehicle, This represents the confidence score of the detection box.

[0110] The training process of the aforementioned lightweight convolutional neural network specifically includes the following steps:

[0111] Step T1: Construct a multi-scene hybrid augmentation dataset. Specifically: To improve the model's generalization ability in complex traffic environments, a dataset of intersection vehicle images covering multiple scenarios, including sunny days, rain / fog, nighttime, and backlighting, is first constructed. Before inputting the images into the network, Mosaic data augmentation technology is used to preprocess the images: four training images are randomly selected, cropped, scaled, color gamut transformed (HSV), and stitched together to reconstruct a single image. New samples of pixels. This method greatly enriches the background of the detected target, especially increasing the frequency of small targets (such as vehicles queuing in the distance) in the samples, which helps to improve the model's sensitivity to the recognition of the "farthest vehicle" in step S1.

[0112] Step T2: Design the multi-task decoupling loss function. Specifically: Considering the characteristics of the network decoupling head, the traditional hybrid loss calculation method is abandoned during training. Instead, separate loss functions are designed for classification and regression tasks, and weights are updated through a multi-task joint optimization strategy. Total loss function. The definition is as follows:

[0113] ;

[0114] In the formula, For the total loss function, For classifying losses, To regress the loss, For confidence loss, , , These are the weight coefficients for the classification loss, regression loss, and confidence loss, respectively.

[0115] in:

[0116] Classification loss ( A binary cross-entropy loss is used to optimize the predicted probability of the "classification branch" to ensure that the model can accurately distinguish between categories such as cars, buses and trucks.

[0117] Regression loss ( For the coordinates output by the "regression branch", CIoU loss is used. Compared with ordinary IoU, CIoU additionally considers the consistency of the center distance and aspect ratio between the predicted box and the ground truth box, which can accelerate convergence and improve the vehicle queue length in step S1. The measurement accuracy. The formula is expressed as:

[0118] ;

[0119] In the formula, For the CIoU loss in the regression loss, The intersection-union ratio (IU) of the predicted bounding box and the ground truth bounding box. These represent the center points of the predicted bounding box and the center points of the ground truth bounding box, respectively. To balance the weight parameters, For Euclidean distance, The diagonal distance of the minimum closure region. Used to measure aspect ratio consistency.

[0120] Confidence loss ( This also employs BCE loss, specifically designed to monitor whether the target exists at the current feature point. The training of this part directly determines the confidence score in step S3. The reliability of the model, i.e., forcing the model to output close to the truth when it detects a real object. of The output in the background area is close to of .

[0121] Step T3: Cosine Annealing-based optimization iteration. Specifically: During the initialization phase, pre-trained weights are loaded onto the large-scale COCO dataset to accelerate the convergence of the feature extraction network. The training optimizer is SGD (Stochastic Gradient Descent), and the initial learning rate is set to... The momentum parameter is During training, a cosine annealing strategy is used to dynamically adjust the learning rate, causing it to decrease according to a cosine function curve with each training epoch. This strategy enables the model to rapidly decrease its learning rate in the early stages of training to find the optimal solution region, and then make minor parameter adjustments in the later stages of training to escape local minima, thereby obtaining more robust weight parameters.

[0122] Step T4: Reparameterization and quantization for edge computing. Specifically: After training, to adapt to the computing power constraints of the edge computing module, the following post-processing is performed:

[0123] Reparameterization: This refers to the parameterization of parameters in the backbone network during the training phase. Convolution and The convolutional branches are merged into a single convolutional structure, reducing the GPU memory usage during inference without sacrificing accuracy.

[0124] PTQ quantization: Using the TensorRT toolchain, the model weights are quantized from 32-bit floating-point numbers (FP32) to 8-bit integers (INT8). By calibrating the dataset and statistically analyzing the activation value distribution of each layer of the network, the optimal quantization threshold is calculated, thereby compressing the model size to a fraction of its original size. To ensure that embedded devices at intersections can achieve The above refers to real-time inference speed.

[0125] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0126] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A traffic signal lamp intelligent remote control method based on multi-source data fusion, characterized in that, include: Traffic flow data at intersections is collected in real time through intersection sensor networks and internet data sources; Establish a unified Cartesian coordinate system and an event-triggered snapshot mechanism for intersections, perform spatiotemporal alignment processing, and spatially map the traffic flow data to the same intersection physical model, and synchronize it in time based on the start time of the signal control cycle; Based on traffic flow data processed by spatiotemporal alignment, the basic green light duration to meet the current traffic demand of the intersection is calculated by using visual confidence weight coefficient and radar confidence weight coefficient, combined with lane saturation dissipation speed and maximum capacity. A duration correction model is constructed by introducing a downstream backpressure suppression factor and a historical residual feedback compensation term to correct the basic green light duration and generate a corrected green light execution duration. The downstream backpressure suppression factor is used to reduce the release time when there is downstream congestion, and the historical residual feedback compensation term is used to compensate for the residual queue in the previous cycle. The cloud server encapsulates the revised green light execution duration into a control command with a lifecycle timestamp and sends it to the mobile intelligent traffic signal light; the mobile intelligent traffic signal light executes the command after verifying its timeliness, and collects residual queuing data at the end of the cycle and sends it back to the cloud for use in revising the scheme for the next cycle. the traffic flow data includes real-time queue length of each lane , vehicle dynamic arrival rate , vehicle departure speed , and real-time saturation of downstream road segments ; The formula for calculating the basic green light duration is as follows: ； In the formula, This indicates the control cycle that is about to be executed. Based on the basic green light duration, The lane saturation dissipation velocity, To maximize the capacity of the lanes, This is the visual reliability weighting coefficient. This is the radar confidence weighting coefficient. Based on the basic safety redundancy factor, In order to match the current control cycle Historical average green light duration for items with the same date attribute and within the same statistical window; The lane saturation dissipation speed The update is performed using an exponentially weighted moving average algorithm, and the update formula is as follows: ； In the formula, The vehicle departure speed for the current cycle. To update the coefficients, Minimum flow rate threshold; The expression for the duration correction model is: ； In the formula, To adjust the green light execution time, This is the downstream back pressure suppression factor term. The sensitivity coefficient, The congestion trigger threshold, For historical residual feedback compensation items, For feedback gain coefficient, This represents the residual queue length at the end of the previous cycle.

2. The intelligent remote control method for traffic lights based on multi-source data fusion according to claim 1, characterized in that, The real-time queue length of each lane By acquiring key frames of surveillance video through high-definition cameras, constructing an inverse perspective transformation matrix to map the image pixel coordinates to road world coordinates, and identifying the Euclidean distance between the farthest vehicle's rear end and the stop line while waiting at a red light; The vehicle dynamic arrival rate The system tracks the trajectory of moving targets entering the field of view using millimeter-wave radar, counts the number of vehicles whose velocity vectors point towards the intersection within a set time window, and calculates the average instantaneous speed of the target crossing the stop line during the green light phase to obtain the vehicle's departure speed. ; The real-time saturation of the downstream road section By separately obtaining the saturation of the congestion delay index conversion of Internet map services Saturation calculated by the geomagnetic induction coil at the downstream intersection Then take the maximum of the two to obtain, that is .

3. The intelligent remote control method for traffic lights based on multi-source data fusion according to claim 2, characterized in that, The spatiotemporal alignment process includes: projecting the visually recognized vehicle coordinates and the radar target coordinates transformed by a rotation and translation matrix onto a unified Cartesian coordinate system using a homography matrix, and performing spatial position matching; and controlling the signal cycle. The start time Open time window as a benchmark Take the arithmetic mean of the radar and visual data within the window, and correlate it with the distance. The most recent valid internet data This is used to ensure the synchronization of traffic flow data over time.

4. The intelligent remote control method for traffic lights based on multi-source data fusion according to claim 1, characterized in that, The visual reliability weight coefficient The radar confidence weighting coefficient is the arithmetic mean of the confidence scores of the queued vehicles identified in the current period. The coefficient of determination for the linear regression of the distance and time trajectory of the radar-tracked target within the current period. The root mean square value.

5. The intelligent remote control method for traffic lights based on multi-source data fusion according to claim 1, characterized in that, The validity period of the verification instruction includes: a timestamp generated by the cloud server containing the instruction's lifecycle. and effective window The mobile intelligent traffic light receives the control message and verifies the current time. The difference: If the difference is less than Then the calculated green light execution time will be executed. If the difference is greater than or equal to If the command is not executed, the instruction will be discarded and the preset minimum green light time will be executed.

6. The intelligent remote control method for traffic lights based on multi-source data fusion according to claim 1, characterized in that, The proposed scheme for correcting the next cycle includes: at the end of the current cycle, the mobile intelligent traffic light uses a visual sensor to calculate the remaining queue length. Record the actual green light duration. It is then uploaded to the cloud to update the historical residual feedback compensation items of the next cycle duration correction model.

7. A traffic signal intelligent remote control system based on multi-source data fusion, used to execute the traffic signal intelligent remote control method based on multi-source data fusion as described in any one of claims 1 to 6, characterized in that, include: The multi-source sensing module is used to collect real-time queue length, vehicle dynamic arrival rate, vehicle departure speed and real-time saturation of downstream road sections for each lane through intersection sensor network and Internet data source. The spatiotemporal alignment module is used to establish a unified coordinate system for intersections and an event-triggered snapshot mechanism, which maps the collected heterogeneous traffic flow data to spatial location and synchronizes it with the time reference, thereby achieving multi-source data fusion. The intelligent computing module is used to calculate the basic green light duration based on the fused data, and to construct a duration correction model by combining the downstream back pressure suppression factor and the historical residual feedback compensation term, so as to dynamically generate the final green light execution duration. The remote control module is used to generate and issue control commands containing lifecycle timestamps, drive the traffic lights to execute after verifying the timeliness of the commands, and collect residual queuing data after execution for closed-loop feedback.

Images