Road traffic sign device and control method thereof

Abstract

The present application belongs to the technical field of traffic control system, and particularly relates to a road traffic sign device and a control method thereof. The method comprises collecting multi-source perception data of a target road section; extracting multi-dimensional feature parameters from the multi-source perception data, assigning fusion weights according to the quality differences of the multi-dimensional perception data, and obtaining comprehensive state data through weighted fusion; inputting the comprehensive state data into a decision model based on machine learning, outputting a control strategy corresponding to the current traffic state, controlling a display unit to perform display output and execute a steering operation to a target steering angle; obtaining traffic state change information after the display output, generating an effect evaluation result, and adjusting and updating parameters of the decision model according to the effect evaluation result. The present application realizes adaptive perception fusion and closed-loop decision optimization of the road traffic sign device under complex environmental conditions.

Description

Technical Field

[0001] This invention belongs to the field of traffic control system technology, specifically relating to a road traffic sign device and its control method. Background Technology

[0002] Road traffic signs are the infrastructure of road traffic management systems, used to convey traffic rules, road condition information, and guidance to road users. Traditional road traffic signs are mainly passive and static, with their displayed content remaining unchanged after installation, unable to adapt to dynamic changes in road traffic conditions. Therefore, some existing technologies have emerged that use sensors to perceive environmental information and adjust the sign's status accordingly.

[0003] The common deficiency of existing solutions lies in the lack of a complete closed-loop mechanism for signage devices to autonomously make decisions based on their own perceived information and continuously optimize their decision-making capabilities based on the implementation effects. Specifically, existing solutions typically respond passively based on only a single dimension of perceived information or preset rules; when multiple sensors are configured, the fusion weights of multi-source information are fixed before deployment and cannot be adaptively adjusted according to the dynamic changes in the real-time data quality of each sensor. When the data quality of some sensors deteriorates due to environmental interference, unreliable perceived information still participates in fusion with fixed weights, affecting the reliability of the overall perception result; whether the control strategy has produced the expected improvement in traffic conditions cannot be automatically evaluated, and the evaluation result cannot be fed back to the decision-making model to optimize subsequent decisions. The decision-making capability of the device does not improve with the increase of operating time. These shortcomings make it difficult for existing signage devices to continuously and reliably output control decisions that conform to the actual traffic conditions in complex and ever-changing road traffic environments.

[0004] To address the aforementioned problems, this invention provides a road traffic sign device and its control method. By performing adaptive fusion of multi-source sensing data, decision-making calculations, and control effect feedback optimization locally on the device, the sign device possesses independent and continuously optimized closed-loop decision-making capabilities. Summary of the Invention

[0005] In view of the above-mentioned shortcomings in the prior art, the present invention provides a road traffic sign device and its control method to solve the problems in the background art.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A method for controlling a road traffic sign device includes the following steps: Step 1: Collect multi-source sensing data for the target road segment; the multi-source sensing data includes at least two different types of environmental parameters; Step 2: Extract multi-dimensional feature parameters representing the traffic status of the target road segment from the multi-source sensing data. Based on the quality differences of the sensing data corresponding to each dimension feature parameter, assign corresponding fusion weights to the feature parameters of each dimension, and perform weighted fusion of the feature parameters of each dimension to obtain comprehensive status data. Step 3: Input the comprehensive status data into the decision model built on machine learning. The decision model outputs a control strategy corresponding to the current traffic status. The control strategy includes the display content, display method and steering angle of the road traffic sign device. Step 4: Generate drive control signals according to the control strategy. The drive control signals control the display unit to display and output according to the specified display content and display mode, and at the same time control the road traffic sign device to turn to the specified turning angle. Step 5: Obtain traffic status change information after display output, generate effect evaluation results, and adjust and update the parameters of the decision model based on the effect evaluation results.

[0007] Furthermore, the method for obtaining the quality difference of the sensing data in step 2 is as follows: obtain the signal quality index, the consistency index with other sensing data in the same period, and the stability index within a preset time window for each sensing data; multiply the above three types of indicators after normalization to determine the quality evaluation value of the sensing data in each dimension.

[0008] Furthermore, the fusion weights in step 2 are calculated according to the following weight formula: in, For the first Each dimension of feature parameters Moment-wise fusion weights For the first Data perceived from multiple dimensions Quality assessment value at any moment For the first The deviation of the perceived data quality assessment value from its historical mean in each dimension. The preset deviation penalty coefficient, This represents the total number of dimensional feature parameters involved in the fusion.

[0009] Furthermore, the comprehensive status data is calculated according to the following formula: in, for Comprehensive state data at any given time. For the first Each dimension of feature parameters The eigenvector at time step 1.

[0010] Furthermore, the effect evaluation result in step 5 is to collect the changes of each traffic state evaluation index before and after the implementation of the control strategy; after normalizing each change, the average value is taken to obtain the effect evaluation value.

[0011] Furthermore, in step 5, the adjustment of the decision model parameters involves using the performance evaluation value as a product factor of the gradient of the output strategy relative to the model parameters to determine the adjustment amount of the model parameters. in, For decision-making models in The parameter adjustment amount at any given time. for The effect evaluation value at any moment, To integrate the state data of the decision model at the previous time step The output strategy is the gradient with respect to the model parameters.

[0012] Furthermore, the multi-dimensional feature parameters in step 2 include at least two of the following parameters: parameters characterizing the density of traffic participants in the target road segment, parameters representing the speed of traffic participants, parameters representing traffic flow stability, parameters representing ambient lighting conditions, parameters representing visibility conditions, parameters representing meteorological conditions, and parameters representing road surface adhesion status.

[0013] Furthermore, the decision model in step 3 adopts an end-to-end decision-making approach, mapping comprehensive state data to control strategies; the control strategies include display content, display mode, display brightness, and turning angle.

[0014] A road traffic sign device, comprising: A road pole, comprising a vertical pole and a horizontal pole, wherein the lower end of the vertical pole is installed on the road surface foundation, and the horizontal pole is set at the top of the vertical pole. The horizontal pole is rotatably connected to the top of the vertical pole, and the horizontal pole is capable of rotating in the horizontal plane relative to the vertical pole. A steering drive mechanism is located at the bottom end of the vertical rod, and the power output end of the steering drive mechanism is connected to one end of the horizontal rod through a transmission mechanism. Multiple environmental sensing sensors are installed on the poles; The data preprocessing module communicates with multiple environmental sensing sensors to perform time synchronization processing and spatial coordinate registration on the environmental information collected by the multiple environmental sensing sensors, and generates a multi-source sensing dataset. The multi-source information fusion module communicates with the data preprocessing module, extracts multi-dimensional feature parameters reflecting the traffic status of the target road segment from the multi-source sensing data, assigns corresponding fusion weights to the feature parameters of each dimension according to the quality differences of the sensing data corresponding to each dimension feature parameter, and performs weighted fusion of the feature parameters of each dimension to obtain comprehensive status data. The edge intelligent decision-making module is communicatively connected to the multi-source information fusion module and has a built-in machine learning-based decision model. The decision model outputs a control strategy corresponding to the current traffic state based on comprehensive state data. The control strategy also includes the steering angle for the crossbar. The display unit, located on the crossbar and connected to the edge intelligent decision-making module, is a variable information display panel that adjusts the display content and display mode according to the control strategy. The steering control module is connected to the edge intelligent decision module and the steering drive mechanism. It generates a steering drive signal according to the control strategy. The steering drive signal controls the steering drive mechanism to drive the crossbar to rotate to the target steering angle. The feedback optimization module is connected to the multi-source information fusion module and the edge intelligent decision-making module. It acquires the traffic state change information of the target road segment after the display unit displays the output according to the control strategy, generates the effect evaluation result based on the traffic state change information, and adjusts and updates the parameters of the decision model based on the effect evaluation result. The power supply module is electrically connected to the above modules.

[0015] Furthermore, the steering drive mechanism includes a drive motor and a reducer; the output shaft of the drive motor is connected to the input shaft of the reducer, the output shaft of the reducer is connected to one end of the crossbar, one end of the crossbar is provided with a rotating engagement part, the rotating engagement part is engaged with a rotating connecting seat, and the rotating connecting seat is located at the top of the vertical rod; the multiple environmental perception sensors are at least two of the following: visual sensor, radar sensor, meteorological sensor, and road condition sensor.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention collects multi-source sensing data of a target road segment by setting up multiple environmental sensing sensors, and adaptively allocates fusion weights locally on the device according to the actual data quality differences of each sensor. This allows the fusion weights to be adjusted in real time in response to the dynamic changes in sensor data quality. When the data quality of a sensor deteriorates due to environmental interference, the fusion weight corresponding to that sensor automatically decreases, and the contribution of unreliable sensing information to the fusion result is reduced accordingly, thereby improving the reliability of the comprehensive status data under complex environmental conditions. 2. Deploying the decision-making model locally on the road traffic sign device enables the device to independently complete the entire computation process from sensing data input to control strategy output without relying on external network connections. This setup eliminates the impact of data transmission delays and network interruptions on the real-time performance of the device's decisions, allowing the sign device to operate normally in areas with weak network coverage or under conditions of communication link failure, thus improving the device's operational reliability in various deployment scenarios. 3. By acquiring information on changes in the actual traffic conditions of the target road segment after the implementation of the control strategy, an effect evaluation result is generated. This evaluation result is used as the basis for online adjustment of the decision model parameters, enabling the device's decision-making capability to be continuously optimized based on the actual control effect. This closed-loop mechanism allows the identification device to gradually adjust its decision-making tendency according to historical control effects under different traffic scenarios, time periods, and weather conditions, improving the device's long-term adaptability to the characteristics of the deployment scenario. Attached Figure Description

[0017] Figure 1 This is a step diagram of a road traffic sign device and its control method according to the present invention; Detailed Implementation To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0018] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual images. They should not be construed as limiting the scope of this application. To better illustrate the embodiments of the present invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0019] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present application. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0020] In the description of this invention, unless otherwise explicitly specified and limited, the term "connection" or similar designation indicating a connection between components should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0021] Example 1: like Figure 1 As shown, the present invention provides a control method for a road traffic sign device, comprising the following steps: Step 1: Collect environmental information of the target road section by using multiple environmental sensing sensors installed on the road traffic sign device to obtain multi-source sensing data including at least two different types of environmental parameters. Perform time synchronization processing and spatial coordinate registration on the multi-source sensing data to generate multi-source sensing data. Step 2: Extract multi-dimensional feature parameters representing the traffic state of the target road segment from the multi-source sensing data. Based on the quality differences of the sensing data corresponding to each feature parameter, evaluate the quality of the sensing data corresponding to each feature parameter, obtaining the quality assessment value of each dimension and its deviation from its historical quality level. Determine the fusion weight of each feature parameter based on its quality assessment value and deviation. A higher quality assessment value results in a larger fusion weight for that feature parameter, and a larger deviation from its historical level results in a greater suppression of the fusion weight. Weighted fusion of the feature parameters according to their determined fusion weights generates comprehensive state data. Furthermore, the multi-dimensional feature parameters include at least two of the following parameters: parameters characterizing the density of traffic participants in the target road segment, parameters representing the speed of traffic participants, parameters representing traffic flow stability, parameters representing ambient lighting conditions, parameters representing visibility conditions, parameters representing meteorological conditions, and parameters representing road surface adhesion status.

[0022] The method for obtaining the quality difference of the sensing data is as follows: obtain the signal quality index, the consistency index with other sensing data in the same period, and the stability index within a preset time window for each sensing data; after normalizing the above three types of indicators, multiply them to determine the quality assessment value of the sensing data in each dimension.

[0023] Furthermore, the fusion weights are calculated using the following weighting formula: in, For the first Each dimension of feature parameters Moment-wise fusion weights For the first Data perceived from multiple dimensions Quality assessment value at any moment For the first The deviation of the perceived data quality assessment value from its historical mean in each dimension. The preset deviation penalty coefficient, This represents the total number of dimensional feature parameters involved in the fusion.

[0024] Specifically, For the first Data perceived from multiple dimensions The quality assessment value at time is solved in the following way, sequentially... Data perceived from multiple dimensions Signal quality metrics, consistency metrics, and stability metrics at any given moment; in, As a signal quality indicator, As a consistency indicator, This is a stability indicator.

[0025] Signal quality metrics are used to characterize the quality of the sensor's raw signal; they are obtained in real-time by the data preprocessing module from the raw signal or data output by the sensor. For visual sensors, the data preprocessing module first calculates three sub-metrics for the current frame image: sharpness, contrast, and signal-to-noise ratio. Sharpness is calculated using the Brenner gradient function, with the following formula: The contrast ratio is calculated using the root mean square contrast ratio, and the formula is as follows: Signal-to-noise ratio (SNR) is the ratio of signal power to noise power. The signal power is calculated as the variance of the pixel grayscale values ​​in the effective area of ​​the image, and the noise power is calculated as the variance of the pixel grayscale values ​​in the uniform background area of ​​the image. These are mapped to the [0,1] interval using an S-shaped function. After normalizing the above three sub-indices to the [0,1] interval, the geometric mean of the three is taken as the signal quality index of the visual sensor. For radar sensors, the data preprocessing module calculates the signal-to-noise ratio based on the ratio of the radar echo signal power to the receiver noise power, and combines the echo signal strength to comprehensively determine the signal quality index. For meteorological sensors and road condition sensors, the amplitude stability (reciprocal of standard deviation) and zero drift of the output signal are calculated and normalized to [0,1].

[0026] The signal quality indicators of all types of sensors are normalized to the [0,1] range, with larger values ​​indicating better signal quality.

[0027] The consistency index is used to characterize the degree of logical consistency between the data collected by a sensor and the data collected by other sensors during the same period. This index is calculated by the multi-source information fusion module after comprehensively comparing the output data of each sensor at the same time: when the data of a certain sensor deviates significantly from the data of other sensors in terms of physical logic (for example, a camera detects that the traffic density in a certain lane is higher than a preset density threshold, while the radar detects that the average vehicle speed in that lane is higher than a preset speed threshold, which constitutes a logical contradiction in traffic flow theory), the consistency index of that sensor will be reduced accordingly.

[0028] Taking visual and radar data as an example: if the traffic density detected by vision is... The average vehicle speed detected by radar is Then the theoretical flow rate is defined. Compare the actual flow rate with the expected flow rate obtained from historical statistics. Comparison, expected traffic Obtained by taking the arithmetic mean of historical traffic data for the same period over the past 24 hours; Consistency index: in This is a preset sensitivity coefficient. The larger the deviation, the lower the sensitivity coefficient. The closer to 0, the closer to 1. Similar principles are used to calculate the consistency of other sensor combinations.

[0029] The consistency index is normalized to the [0,1] range. The larger the value, the higher the logical consistency between the sensor data and other sensor data.

[0030] The stability index characterizes the degree of fluctuation in data quality of the sensor within a preset time window. This index is calculated by the data preprocessing module after statistical analysis of the data quality of a single sensor within the preset time window. It is based on the variance of the signal quality index within this time window, and its calculation formula is as follows: ;in To adjust the parameters, the stability index approaches 1 when the standard deviation approaches 0. When the variance of the signal quality index is small within this time window, it indicates stable data quality, and the stability index value is high; when the variance is large, it indicates drastic fluctuations in data quality, and the stability index value is low. The stability index is normalized to the [0,1] interval.

[0031] In summary, after normalizing the above three types of indicators and multiplying them together, we obtain the first... Data perceived from multiple dimensions Quality assessment value at any moment Since all three indices have been normalized to the [0,1] interval, their product... It also lies in the [0,1] interval. The larger the value, the higher the overall quality and the stronger the credibility of the perceived data in that dimension.

[0032] Further calculation of the first Deviation of the perceived data quality assessment value from its historical mean in each dimension Regarding the first Each dimension of perceived data is used to calculate the historical mean based on the historical quality assessment value sequence of that dimension. In this embodiment, Calculated using an exponentially weighted moving average (a fundamental algorithm in existing signal processing and statistics): in, This is the historical average. The historical mean is not limited to the above-mentioned exponentially weighted moving average algorithm; other equivalent methods such as sliding window mean or long-term statistical mean can also be used. That is, the historical mean at the current moment is obtained by weighting the historical mean at the previous moment and the quality assessment value at the previous moment according to the preset forgetting factor, so that the historical data closer to the current moment has a greater impact on the mean.

[0033] Then calculate the deviation: The sign and absolute value of the deviation represent the direction and degree of deviation of the current data quality from the historical norm; when When the value is negative and large in absolute terms, it indicates that the current data quality is significantly worse than historical levels, and its fusion weight should be effectively suppressed; when When the value is close to zero or positive, it indicates that the current data quality is at or above the historical normal level, and no additional suppression is required; it can participate in the fusion process normally.

[0034] Preset deviation penalty coefficient This is used to control the suppression strength of the bias on the fusion weights, and its value typically ranges from 0.5 to 5.0, being a real number greater than or equal to zero. When When the bias does not have any inhibitory effect on the fusion weights, the fusion weights are determined solely by the quality assessment value. In this case, the system degenerates into a traditional data quality-weighted fusion, which does not include a bias suppression mechanism. At that time, the exponentially decaying term As the absolute value of the deviation increases, it decreases monotonically, and the degree to which the fusion weight of the perceived data in this dimension is suppressed increases accordingly.

[0035] Preset deviation penalty coefficient The value is preset based on the characteristics of the sensor type and application scenario, or it can be determined through offline simulation optimization: Perform a parameter scan on the variables and select the ones that maximize the reliability of the fusion results under various operating conditions. Value. In this embodiment, The basis for this is: selecting sensor data under typical operating conditions such as sunny days, rain / fog, and backlighting, to... The parameters were scanned within the range of 0.5 to 5.0. With the goal of minimizing the deviation between the integrated state data after fusion under various working conditions and the actual traffic conditions, the optimal value was determined to be 2.0.

[0036] Furthermore, the comprehensive status data is calculated according to the following formula: in, for Comprehensive state data at any given time. For the first Each dimension of feature parameters The eigenvector at time step 1.

[0037] Specifically, the multi-source information fusion model extracts features from various types of perceived data to obtain feature parameters for each dimension. For image data, a deep learning-based target detection algorithm is used to identify vehicle targets in the image. Lane occupancy rate is obtained by comparing the pixel area occupied by vehicle targets within the detection area with the total pixel area of ​​the detection area; traffic density is obtained by dividing the total number of vehicles within the detection area by the actual road length corresponding to the detection area; thus, the feature value vector of the first dimension feature parameter is generated. , It includes two characteristic values: lane occupancy rate and traffic density.

[0038] For radar data, a Kalman filter algorithm is used to track the detected vehicles. The average vehicle speed is obtained by taking the arithmetic mean of the velocity components along the road direction for each vehicle within the detection range; the average headway is obtained by taking the arithmetic mean of the time differences between adjacent vehicles in the same lane passing the same reference section. This generates the eigenvalue vector of the first-dimensional feature parameters. , It includes two characteristic values: average vehicle speed and average headway.

[0039] For meteorological data, the measured values ​​of ambient light intensity are quantified and classified according to a preset light intensity classification standard to obtain light intensity level parameters; the measured values ​​of visibility are quantified and classified according to a preset visibility classification standard to obtain visibility level parameters; meteorological condition categories are coded to obtain meteorological type coding parameters; and thus, a feature value vector of the third-dimensional feature parameters is generated. , It includes three feature values: illumination level, visibility level, and weather type code.

[0040] For pavement condition data, based on the measured pavement moisture content, the pavement condition category is identified according to a preset moisture content threshold range, pavement condition coding parameters are obtained, and the estimated pavement adhesion coefficient is obtained by mapping the pavement condition coding parameters, thereby generating the feature value vector of the fourth dimension feature parameters. , It includes two feature values: pavement condition code and pavement adhesion coefficient estimate.

[0041] Step 3: Input the comprehensive status data into the decision model built on machine learning. The decision model outputs a control strategy corresponding to the current traffic status. The control strategy includes the display content, display method and steering angle of the road traffic sign device. Furthermore, the decision-making model adopts an end-to-end decision-making approach, mapping comprehensive state data to control strategies; the control strategies include display content, display mode, display brightness, and turning angle.

[0042] Specifically, the decision-making model is a pre-trained lightweight neural network model. The structure of this lightweight neural network model is as follows: the number of nodes in the input layer corresponds to the number of dimensions in the comprehensive state data; the hidden layer has two layers, each containing 64 nodes, with the ReLU activation function; the number of nodes in the output layer corresponds to the number of all possible control strategies; the output layer uses the Softmax activation function, and the output is the probability distribution of each strategy. Before being deployed locally to the device, the decision-making model was trained offline using a training dataset containing various traffic scenarios. The offline training dataset contains 5000 labeled samples, covering various weather conditions such as sunny, rainy, foggy, and nighttime, as well as various traffic states such as smooth traffic, congestion, and accidents. The samples are evenly distributed according to scenario type. Specific annotations are as follows: when traffic density is less than 15 veh / km and average vehicle speed is higher than 90% of the speed limit, the annotation is "Speed ​​limit displayed + constant light + normal brightness + straight ahead"; when visibility is less than 100m and road surface adhesion coefficient is less than 0.4, the annotation is "Slippery and slow + flashing + high brightness + 15° left deviation"; other scenarios follow the same pattern. All annotations were independently annotated by at least three traffic engineering experts, achieving majority consensus. The training loss function is a weighted sum of cross-entropy loss and policy smoothing loss (weight ratio 0.8:0.2), with safety-critical categories (such as accident warnings) additionally multiplied by a weight coefficient of 2.0 in the cross-entropy loss. The optimizer uses Adam with an initial learning rate of 0.001, a batch size of 64, and 200 training epochs. The early stopping mechanism is triggered when the validation set loss does not decrease for 20 consecutive epochs.

[0043] The decision model performs inference calculations locally on the device based on the input comprehensive state data, and outputs a control strategy corresponding to the current traffic state. The control strategy includes at least information specifying the type of content displayed by the road traffic sign device, information specifying the display mode, information specifying the display brightness level, and information specifying the crossbar turning angle. All of the above output items are natural outputs of the decision model's inference calculations, and are not obtained by looking up a table based on a manually set risk level.

[0044] The decision-making model employs an end-to-end approach, directly mapping comprehensive state data to control strategies. To ensure the rationality of the model's output control strategies, the training dataset used in the offline training phase includes various training samples labeled with the optimal control strategies based on expert experience, and the training loss function is weighted differently according to the safety priorities of different traffic scenarios. Through this training method, after deployment, the decision-making model can directly output a control strategy matching the current traffic state based on any input comprehensive state data, without needing to set risk level thresholds during operation.

[0045] Optionally, to facilitate understanding of the decision-making logic of the decision model, the output strategies of the model can be qualitatively classified as follows: When the feature value reflecting traffic flow density in the comprehensive state data is lower than the empirical threshold of the smooth flow state in the training dataset, the decision model tends to output a control strategy of displaying a regular speed limit, constant brightness, and normal brightness; when the feature value reflecting average vehicle speed in the comprehensive state data is lower than the empirical threshold of the congested state in the training dataset, and the feature value reflecting headway decreases, the decision model tends to output a control strategy of displaying congestion warnings; when the feature value reflecting visibility in the comprehensive state data is lower than the empirical threshold of the low visibility state in the training dataset, or the feature value reflecting road surface adhesion is lower than the empirical threshold of the slippery road surface state in the training dataset, the decision model tends to output a control strategy of displaying safety warning information, flashing mode, and high brightness. The above empirical thresholds are obtained statistically from the sample distribution in the training dataset and are not fixed thresholds preset during device operation. The different control strategies output by the decision model differ in display content, display method, display brightness, and steering angle; these differences are jointly determined by the model parameters and input data.

[0046] Furthermore, the offline training of the decision-making model is explained: the decision-making model is trained offline using supervised learning. The training dataset consists of multiple traffic scenario samples, each sample including comprehensive state data and a corresponding labeled control strategy. The labeled control strategies are annotated by traffic engineering experts according to road traffic design specifications and safe passage principles. The training loss function is a weighted combination of cross-entropy loss and policy smoothing loss, where safety-related control strategy categories (such as accident warnings, slippery road surface alerts, etc.) are given higher weight coefficients in the loss function to ensure the model's accuracy in making decisions for safety-critical scenarios. The Adam optimizer is used during training, with a learning rate set to 10. -4 The training epochs are 200, and the early stopping mechanism is triggered in 20 epochs. The trained model parameters are then fixed and deployed to the edge intelligent decision-making module of the road traffic sign device.

[0047] Step 4: Generate drive control signals according to the control strategy. The drive control signals control the display unit to display output according to the specified display content and display mode, and at the same time control the road traffic sign device to turn to the specified turning angle. The drive control signals include at least the display content encoding signal, the display mode control signal, the brightness adjustment signal, and the angle signal.

[0048] Specifically, the control strategy is structured data output by the decision model, which includes at least the following fields: content type field, used to specify the type of information to be displayed by the display unit; mode field, used to specify the working mode of the display unit; brightness field, used to specify the luminous intensity level of the display unit; and angle field, used to specify the target turning angle of the crossbar.

[0049] The method for generating drive control signals according to the control strategy is as follows: the content type field is mapped to display content data that the display unit can recognize. This mapping is achieved through a preset content-encoding mapping table, which is stored in the storage module of the road traffic sign device and contains dot matrix encoding data corresponding to speed limit values, lane indicator arrows, text warning information, and graphic symbols. The mode field is mapped to the control level used to control the working mode of the display unit. The brightness field is mapped to a pulse width modulation signal with a corresponding duty cycle, and the duty cycle is positively correlated with the brightness level. The angle field is mapped to a steering drive signal, which is used to turn the road traffic sign device to a specified steering angle.

[0050] The drive control signals are sent to the display unit and steering control module of the road traffic sign device, respectively. The display unit decodes and displays the corresponding content according to the received signals, and the steering control module controls the steering drive mechanism to drive the crossbar to rotate to the specified steering angle according to the received signals.

[0051] Step 5: Obtain the traffic state change information after the display output, generate the effect evaluation result, and adjust and update the parameters of the decision model based on the effect evaluation result. The effect evaluation result is obtained by collecting the changes of each traffic state evaluation index before and after the control strategy is implemented; after normalizing each change, the average value is taken to obtain the effect evaluation value.

[0052] Furthermore, the performance evaluation value is calculated according to the following formula: in, For the first The change of each traffic condition evaluation index after the implementation of the control strategy relative to before its implementation. For the first Reference values ​​for each traffic condition evaluation indicator. This represents the total number of traffic condition evaluation indicators.

[0053] Specifically, the evaluation results are obtained by prioritizing the determination of traffic state evaluation indicators for assessing the effectiveness of the control strategy. Each traffic state evaluation indicator includes at least one of the following: average vehicle speed, traffic density, vehicle speed standard deviation, and average headway within the target road segment. Each evaluation indicator is calculated using data collected by environmental perception sensors: average vehicle speed and vehicle speed standard deviation are calculated statistically from the instantaneous speeds of each vehicle detected by radar sensors; traffic density is obtained by dividing the number of vehicles detected by visual sensors by the road segment length; and average headway is obtained by averaging the time differences between adjacent vehicles passing the same cross-section as detected by radar sensors.

[0054] The changes in each evaluation indicator before and after the control strategy is executed are collected. "Before execution" refers to the moment before the display unit begins displaying output according to the control strategy. After execution, the output will be displayed after a predetermined time interval. The moment Change Pre-defined time interval The value is set at 60 seconds. It is determined based on the sum of the average reaction time of a driver from receiving visual information to completing the corresponding driving operation in the field of traffic engineering and the time required for the vehicle to travel to the key position downstream of the sign device at the target road speed. This ensures that the control strategy has a sufficient time window to act on the traffic flow and produce observable changes.

[0055] Reference values ​​for each evaluation indicator The reference values ​​are those given in the road design specifications. For average vehicle speed, the reference value is the road design speed for that road segment; for traffic density, the reference value is the standard traffic density corresponding to the lane design capacity of that road segment; for standard deviation of vehicle speed, the reference value is 10% of the design speed of that road segment; for average headway, the reference value is the headway corresponding to the safe following distance at the design speed of that road segment. All of the above reference values ​​are known parameters published in the road design specifications and are determined by the road grade and design speed.

[0056] Furthermore, the adjustment of the decision model parameters involves using the performance evaluation value as a factor of the product of the output strategy and the gradient of the model parameters to determine the adjustment amount of the model parameters. in, For decision-making models in The parameter adjustment amount at any given time. for The effect evaluation value at any moment, To integrate the state data of the decision model at the previous time step The output policy gradient with respect to the model parameters is given. for The learning rate at any given moment.

[0057] Specifically, for The learning rate at each time step is used to control the step size of a single parameter adjustment. It can be set to a fixed value or a decay function that decreases with running time. In this embodiment, The value is a pre-set fixed value of 0.001. This is based on the fact that during the offline training phase, the decision model is trained using the Adam optimizer with a learning rate of 0.001, and the trained model converges stably. Using this learning rate during the online adjustment phase ensures that the parameter adjustment process is consistent with the update scale during the training phase. When the value is 0.001, the magnitude of a single parameter adjustment does not exceed 0.1% of the total variance of the model parameters, thus avoiding drastic fluctuations in the model output due to excessive adjustments in a single instance. It should be noted that... The values ​​mentioned above are only specific implementation methods in this embodiment. In actual deployment, they can be adjusted according to the structural complexity and parameter scale of the decision model. The range is usually controlled within 10 -4 Up to 10 -2 between.

[0058] To integrate the state data of the decision model at the previous time step The gradient of the output policy with respect to the model parameters is calculated by taking the combined state data from the previous time step. The input decision model is used for forward propagation calculations to obtain the policy probability distribution output by the model. Then, the model parameters are adjusted using the probability values ​​of this output strategy. Calculate the partial derivatives to obtain the gradient components of each parameter. The decision model is a neural network model, and the gradient calculation is automatically completed through the neural network backpropagation algorithm. Specifically, it includes: using the policy probability distribution of the model's output layer as the target value, backpropagating layer by layer using the partial derivatives of this target value with respect to the weights of each layer, and calculating the parameter gradients of the hidden layer and the input layer in sequence. The above gradient calculation is executed on the device's local embedded processor, and the time required for a single gradient calculation does not exceed 1.5 times the time required for a single forward inference of the model.

[0059] In actual engineering implementation, the system deploys a neural network inference engine, which supports both forward inference and backward gradient calculation. Specifically, parameter adjustment involves the processor acquiring performance evaluation values ​​within a predetermined time window after the control strategy is executed. Then, read the comprehensive state data from the previous moment from memory. and current model parameters The inference engine is based on Execute forward propagation to obtain output strategy Then, backpropagation is performed to calculate the gradient according to the output strategy. ;Will Multiply by the gradient and then by the learning rate , obtain parameter adjustment amount ;Will Superimposed on current model parameters The parameters are then updated. The updated model parameters are written to the embedded processor's non-volatile memory for use in the next time step's decision-making inference. The entire parameter adjustment process is completed locally on the device, without the need for communication with external devices.

[0060] A road traffic sign device, comprising: A road pole consists of a vertical pole and a horizontal pole. The lower end of the vertical pole is installed on the road surface foundation, and the horizontal pole is set at the top of the vertical pole. The horizontal pole is rotatably connected to the top of the vertical pole and can rotate relative to the vertical pole in the horizontal plane. The steering drive mechanism is located at the bottom of the vertical rod, and the power output end of the steering drive mechanism is connected to one end of the horizontal rod through a transmission mechanism. Multiple environmental sensing sensors are installed on the poles; The data preprocessing module communicates with multiple environmental sensing sensors to perform time synchronization processing and spatial coordinate registration on the environmental information collected by the multiple environmental sensing sensors, generating a multi-source sensing dataset. The multi-source information fusion module is communicatively connected to the data preprocessing module. It extracts multi-dimensional feature parameters reflecting the traffic status of the target road segment from the multi-source sensing data. Based on the quality differences of the sensing data corresponding to each dimension feature parameter, it assigns corresponding fusion weights to each dimension feature parameter and performs weighted fusion on the feature parameters of each dimension to obtain comprehensive status data. In this embodiment, the various environmental sensing sensors include visual sensors (high-definition cameras), radar sensors (millimeter-wave radar), meteorological sensors, and road surface condition sensors.

[0061] The edge intelligent decision-making module communicates with the multi-source information fusion module and has a built-in machine learning-based decision-making model. The decision-making model outputs a control strategy corresponding to the current traffic state based on comprehensive state data. The control strategy also includes information for the steering angle of the crossbar. The display unit, located on the crossbar and connected to the edge intelligent decision-making module, is a variable information display panel that adjusts the display content and display mode according to the control strategy. The steering control module is connected to the edge intelligent decision module and the steering drive mechanism. It generates a steering drive signal according to the control strategy. The steering drive signal controls the steering drive mechanism to drive the crossbar to rotate to the target steering angle. The feedback optimization module is connected to the multi-source information fusion module and the edge intelligent decision-making module. It obtains the traffic state change information of the target road segment after the display unit displays the output according to the control strategy, generates the effect evaluation result based on the traffic state change information, and adjusts and updates the parameters of the decision model based on the effect evaluation result. The power supply module, electrically connected to each module, provides operating power to all modules. In this embodiment, the power input terminal of the power supply module is connected to the municipal power grid, obtaining AC mains power through power lines buried in the road foundation or laid inside the vertical pole. The power supply module internally includes a switching power conversion circuit that converts the AC mains power into the DC operating voltage required by each module, including: DC voltage for the embedded processor and sensors, DC voltage for the communication module, and DC voltage for the display unit and steering drive motor. The output terminal of the power supply module is connected to the data preprocessing module, multi-source information fusion module, edge intelligent decision-making module, display unit, steering control module, and communication module respectively through power lines arranged inside the vertical pole. Each module draws power from its corresponding power line to operate.

[0062] Optionally, the power supply module is also equipped with a backup battery pack, which is connected in parallel with the output of the switching power supply conversion circuit. When the municipal power grid is supplying power normally, the backup battery pack is in a float charging state. When the municipal power grid is interrupted, the backup battery pack automatically switches to a discharging state to provide uninterrupted power supply to each module of the device.

[0063] A communication module, mounted on the vertical pole, is communicatively connected to the edge intelligent decision-making module and is used to exchange status information or control strategy information with other road traffic sign devices. In this embodiment, the communication module adopts a dedicated short-range communication module based on the IEEE 802.11p standard.

[0064] Furthermore, the steering drive mechanism includes a drive motor and a reducer; the output shaft of the drive motor is connected to the input shaft of the reducer, the output shaft of the reducer is connected to one end of the crossbar, one end of the crossbar is provided with a rotating engagement part, the rotating engagement part is engaged with a rotating connecting seat, and the rotating connecting seat is located at the top of the vertical rod; The environmental perception sensor includes at least two of a first type of sensor, a second type of sensor, a third type of sensor, and a fourth type of sensor. The first type of sensor collects traffic image information of the target road segment, the second type of sensor collects the position and motion state information of moving objects within the target road segment, the third type of sensor collects meteorological environmental information of the target road segment, and the fourth type of sensor collects road surface condition information of the target road segment. The environmental perception sensor includes at least two of a visual sensor, a radar sensor, a meteorological sensor, and a road surface condition sensor.

[0065] The above are merely embodiments of the present invention. The circuits, electronic components, and modules involved are all prior art, fully achievable by those skilled in the art, and require no further explanation. The content protected by this application does not involve improvements to the software and methods. Commonly known structures and characteristics in the solutions are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are able to access all prior art in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent.

Claims

1. A control method for a road traffic sign device, characterized in that: Includes the following steps: Step 1: Collect multi-source sensing data for the target road segment; the multi-source sensing data includes at least two different types of environmental parameters; Step 2: Extract multi-dimensional feature parameters representing the traffic status of the target road segment from the multi-source sensing data. Based on the quality differences of the sensing data corresponding to each dimension feature parameter, assign corresponding fusion weights to the feature parameters of each dimension, and perform weighted fusion of the feature parameters of each dimension to obtain comprehensive status data. Step 3: Input the comprehensive status data into the decision model built on machine learning. The decision model outputs a control strategy corresponding to the current traffic status. The control strategy includes the display content, display method and steering angle of the road traffic sign device. Step 4: Generate drive control signals according to the control strategy. The drive control signals control the display unit to display and output according to the specified display content and display mode, and at the same time control the road traffic sign device to turn to the specified turning angle. Step 5: Obtain traffic status change information after display output, generate effect evaluation results, and adjust and update the parameters of the decision model based on the effect evaluation results.

2. The control method for a road traffic sign device as described in claim 1, characterized in that: The method for obtaining the quality difference of the sensing data in step 2 is as follows: obtain the signal quality index, the consistency index with other sensing data in the same period, and the stability index within a preset time window for each sensing data; multiply the signal quality index, consistency index and stability index after normalization to determine the quality evaluation value of the sensing data in each dimension.

3. The control method for a road traffic sign device as described in claim 1, characterized in that: In step 2, the fusion weight is calculated according to the following weight formula: in, For the first Each dimension of feature parameters Moment-wise fusion weights For the first Data perceived from multiple dimensions Quality assessment value at any moment For the first The deviation of the perceived data quality assessment value from its historical mean in each dimension. The preset deviation penalty coefficient, This represents the total number of dimensional feature parameters involved in the fusion.

4. The control method for a road traffic sign device as described in claim 1, characterized in that: The comprehensive status data is calculated according to the following formula: in, for Comprehensive state data at any given time. For the first Each dimension of feature parameters The eigenvector at time step 1.

5. The control method for a road traffic sign device as described in claim 1, characterized in that: The effect evaluation result in step 5 is to collect the changes of each traffic state evaluation index before and after the implementation of the control strategy; after normalizing each change, the average value is taken to obtain the effect evaluation value.

6. The control method for a road traffic sign device as described in claim 1, characterized in that: In step 5, the adjustment of the decision model parameters is determined by using the performance evaluation value as a product factor of the gradient of the output strategy relative to the model parameters, thus determining the adjustment amount of the model parameters. in, For decision-making models in The parameter adjustment amount at any given time. for The effect evaluation value at any moment, To integrate the state data of the decision model at the previous time step The output policy gradient with respect to the model parameters is given. for The learning rate at any given moment.

7. The control method for a road traffic sign device as described in claim 1, characterized in that: The multi-dimensional feature parameters in step 2 include at least two of the following parameters: parameters characterizing the density of traffic participants in the target road segment, parameters representing the speed of traffic participants, parameters representing traffic flow stability, parameters representing ambient lighting conditions, parameters representing visibility conditions, parameters representing meteorological conditions, and parameters representing road surface adhesion status.

8. The control method for a road traffic sign device as described in claim 1, characterized in that: The decision model in step 3 adopts an end-to-end decision-making approach, mapping comprehensive state data to control strategies; the control strategies include display content, display mode, display brightness, and turning angle.

9. A road traffic sign device, used to execute the control method for a road traffic sign device according to any one of claims 1-8, characterized in that: include: A road pole, comprising a vertical pole and a horizontal pole, wherein the lower end of the vertical pole is installed on the road surface foundation, and the horizontal pole is set at the top of the vertical pole. The horizontal pole is rotatably connected to the top of the vertical pole, and the horizontal pole is capable of rotating in the horizontal plane relative to the vertical pole. A steering drive mechanism is located at the bottom end of the vertical rod, and the power output end of the steering drive mechanism is connected to one end of the horizontal rod through a transmission mechanism. Multiple environmental sensing sensors are installed on the poles; The data preprocessing module communicates with multiple environmental sensing sensors to perform time synchronization processing and spatial coordinate registration on the environmental information collected by the multiple environmental sensing sensors, and generates a multi-source sensing dataset. The multi-source information fusion module communicates with the data preprocessing module, extracts multi-dimensional feature parameters reflecting the traffic status of the target road segment from the multi-source sensing data, assigns corresponding fusion weights to the feature parameters of each dimension according to the quality differences of the sensing data corresponding to each dimension feature parameter, and performs weighted fusion of the feature parameters of each dimension to obtain comprehensive status data. The edge intelligent decision-making module is communicatively connected to the multi-source information fusion module and has a built-in machine learning-based decision model. The decision model outputs a control strategy corresponding to the current traffic state based on comprehensive state data. The control strategy also includes the steering angle for the crossbar. The display unit, located on the crossbar and connected to the edge intelligent decision-making module, is a variable information display panel that adjusts the display content and display mode according to the control strategy. The steering control module is connected to the edge intelligent decision module and the steering drive mechanism. It generates a steering drive signal according to the control strategy. The steering drive signal controls the steering drive mechanism to drive the crossbar to rotate to the target steering angle. The feedback optimization module is connected to the multi-source information fusion module and the edge intelligent decision-making module. It acquires the traffic state change information of the target road segment after the display unit displays the output according to the control strategy, generates the effect evaluation result based on the traffic state change information, and adjusts and updates the parameters of the decision model based on the effect evaluation result. The communication module is connected to the edge intelligent decision-making module and is used to exchange status information or control strategy information with other road traffic sign devices; the power supply module is electrically connected to the above modules.

10. A road traffic sign device as described in claim 9, characterized in that: The steering drive mechanism includes a drive motor and a reducer. The output shaft of the drive motor is connected to the input shaft of the reducer. The output shaft of the reducer is connected to one end of the crossbar. One end of the crossbar is provided with a rotating engagement part. The rotating engagement part is engaged with a rotating connecting seat. The rotating connecting seat is located at the top of the vertical rod.

Images