Traffic signal real-time monitoring control system

Abstract

The invention discloses a traffic signal real-time monitoring control system, which comprises monitoring equipment arranged at each traffic intersection in an area, signal lamp controllers in wireless communication connection with the monitoring equipment, and a central control system connected with all signal lamp controllers in the area, and is characterized in that the monitoring equipment is used for shooting the state of each intersection; the central control system is used for shooting traffic lights, preprocessing and feature extraction are carried out on shot images, state data of all intersections are extracted, the information is transmitted to the traffic light controller, and the traffic light controller adopts a reinforcement learning algorithm to adjust the signal conditions of the traffic lights and sends an adjustment scheme to the central control system. The central control system collects the adjustment scheme of each intersection, readjusts the scheme by adopting Nash equilibrium, and sends the adjusted scheme to the signal lamp controller again, and the signal lamp controller performs signal adjustment according to the adjustment scheme sent by the central control system. According to the invention, the traffic flow at the traffic port is improved, and traffic congestion is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of real-time monitoring and control of traffic signals, and in particular to a real-time monitoring and control system for traffic signals. Background Art

[0002] With the acceleration of urbanization and the surge in the number of motor vehicles, traditional traffic signal control systems face severe challenges. Early traffic signal control mostly adopted fixed timing schemes, which achieved signal switching through preset cycles and phase durations. However, its static characteristics made it difficult to adapt to dynamically changing traffic flows, which could easily lead to increased congestion during peak hours or waste of green light resources during off-peak hours. Although adaptive control systems (such as SCOOT and SCATS) were subsequently developed, which use sensors such as geomagnetic coils and cameras to collect traffic data in real time and dynamically adjust timing, their reliance on a single data source and centralized decision-making architecture has significant limitations: the centralized data processing model leads to high communication delays and insufficient response speeds, making it difficult to meet the real-time requirements of complex road networks. In addition, existing systems generally lack the ability to collaboratively optimize multimodal traffic participants (motor vehicles, non-motor vehicles, pedestrians), and are unable to effectively respond to sudden changes in traffic flow caused by emergencies (such as traffic accidents and large-scale events), which can easily cause traffic congestion and affect the normal operation of vehicles.

[0003] The existing patent number is CN111508246A, a traffic signal data transmission method and a traffic signal control system, which include a traffic signal, a traffic flow monitoring module, a wireless communication unit, a traffic flow ratio preset module and a base station center. The output end of the traffic flow monitoring module is connected to the input end of the wireless communication unit. Through the settings of the traffic flow analysis system, the traffic flow ratio preset module and the traffic light intelligent adjustment system, continuous adjustments are made to make the traffic at the intersection smooth, but the adjustment method is not intelligent and cannot make temporary responses according to the existing situation. Each intersection only makes the traffic at the intersection smooth, but the intersections influence and interact with each other, which leads to problems in coordination between the intersections and even contradictions. Therefore, it cannot actually solve the problem of traffic congestion. Summary of the Invention

[0004] In order to at least solve or partially solve the above problems, a traffic signal real-time monitoring and control system is provided to improve the flow of traffic at the intersection, ensure the normal operation of vehicles, and reduce traffic congestion.

[0005] In order to achieve the above object, the present invention provides the following technical solutions:

[0006] The present invention provides a real-time monitoring and control system for traffic signals, comprising monitoring equipment arranged at each traffic intersection in an area, a signal light controller wirelessly connected to the monitoring equipment, and a central control system connected to all signal light controllers in the area, wherein the monitoring equipment comprises a data acquisition module, a data preprocessing module, and a data feature extraction module, wherein the data acquisition module is used to capture the status of each intersection and send the captured image to the data preprocessing module, the data preprocessing module preprocesses the data and sends it to the data feature extraction module, the data feature extraction module extracts features from the data, extracts status data of each intersection, and transmits the information to the signal light controller, the signal light controller adopts a reinforcement learning algorithm to adjust the signal status of the traffic light, and sends the adjusted plan to the central control system, the central control system collects the adjustment plan of each intersection, adopts Nash equilibrium to readjust the plan, and sends the adjusted plan to the signal light controller again, and the signal light controller adjusts the signal according to the adjustment plan sent by the central control system.

[0007] As a preferred technical solution of the present invention, the status data of each intersection includes the saturation of vehicles at each intersection, flow rate, and status of traffic lights.

[0008] As a preferred technical solution of the present invention, the traffic light controller adopts Q reinforcement learning algorithm to adjust the current state of the traffic lights at each intersection at time t and the saturation of the vehicles at the intersection as S t , the strategy adopted at time t is a t , Q(S t , a t ) is the total traffic volume at time t, α is the learning rate, which is adaptively adjusted based on the dynamic traffic load; γ is the discount rate, and the evaluation value of the strategy adopted at time t is r t , then the formula of the reinforcement learning algorithm of Q learning is:

[0009]

[0010] As a preferred technical solution of the present invention, the central control system adopts the Nash equilibrium calculation formula:

[0011]

[0012] in, represents the strategy selected by the i-th traffic light controller; represents the vector of strategies of all traffic light controllers except i; Q i represents the utility level of the i-th traffic light controller; S i represents the strategy space of the i-th traffic light controller; s iRepresents a certain policy of the i-th signal lamp controller, which is a subset of S i ; t represents the time.

[0013] As a preferred technical solution of the present invention, in the monitoring device, feature extraction is performed by cascaded convolution to obtain a feature extraction map. Each time the cascaded convolution cascades N extended convolutional layers, the size of the convolutional kernel is k×k, and the expansion rates are [r1,…,r i ,…,r n .

[0014] As a preferred technical solution of the present invention, the expansion rate r~i~ of the cascaded convolution satisfies r~i~ = 2^i^ to gradually expand the receptive field and extract multi-scale traffic features.

[0015] Define the maximum value of the expansion rate of the i-th layer in the extended convolutional kernel as Ri, and the expansion rate ri ≤ Ri. The calculation formula of Ri is:

[0016]

[0017] And (R1 - 1) + (R2 - 1) < k, when Ri > 0, Ri ≠ Ri+1; the expansion rates ri within each group cannot form a common factor relationship; ri represents the expansion rate of the i-th layer in the extended convolutional kernel, and k is the size of the convolutional kernel.

[0018] Convolution is calculated using batch normalization, and then passed through the activation function Relu. The batch normalization calculation includes the following steps:

[0019] B1: Input the feature maps of the same batch into the batch normalization layer for calculation to obtain the mean value [[ID= thirty-two ]] and variance [[ID= thirty-four ]]

[0020] B2: Normalize all x i , i ∈ 1…m to obtain <00001​​​​​​​​​​​​​​​​​​​​​​​​​​Where m represents the batch capacity; ε represents a very small constant to avoid system errors caused by a denominator of 0; β and γ represent adjustment parameters.

[0024] The adjustment parameters β and γ are both trainable parameters in the neural network. The initial value of γ is 1, and the initial value of β is 0. The calculation formulas of γ and β are as follows: γ′←γ+Δγ; β′←β+Δβ;

[0025] Where y represents the model output image of this training round, represents the labeled image. η represents the learning rate, η = 0.0001.

[0026] As a preferred technical solution of the present invention, the monitoring device sequentially performs grayscale processing, normalization, contrast-limited adaptive histogram equalization, and gamma nonlinear processing to obtain a preprocessed image, wherein the normalization step includes:

[0027] Auxiliary variables are screened using the 3σ principle. Subsequently, maximum and minimum normalization is performed to convert data of different dimensions into data of the same dimension. By mapping the data to [0, 1], the influence of different dimensions is eliminated, making different indicators comparable and improving the convergence speed of the model.

[0028]

[0029] in x t max 、x t min Represent the normalized time series, the maximum value in the data, and the minimum value in the data respectively;

[0030] The normalization comprises the following steps:

[0031] Let X = {x 1 , x 2 ,...,x n} is an image dataset; the Z score normalization method is used to normalize the dimension of the dataset X; the Z score normalization formula is as follows:

[0032] Xnorm=(X-μ) / σ

[0033] Where μ is the mean of x and σ is the standard deviation of x. In this case, x has both positive and negative values, and the mean of x is 0 and the standard deviation is 1.

[0034] Then, each image data in the dataset X is normalized by min-max, and the value of x is remapped to the range of 0 to 255. The min-max normalization formula is as follows:

[0035]

[0036] Where x i ∈X norm , i∈[1,2,...,n].

[0037] As a preferred technical solution of the present invention, the contrast-limited adaptive histogram equalization adopts the CLAHE algorithm to enhance the feature and background contrast of the entire data set, and then uses gamma nonlinearization to perform nonlinear operations on the luminance or tristimulus values ​​in the image to obtain a preprocessed image.

[0038] As a preferred technical solution of the present invention, after the monitoring device obtains the regional image containing the feature, the improved two-dimensional OTSU algorithm is used to segment the regional image to obtain the target area and the background area respectively, and the pixel area S of the target area is calculated. o and with the standard pixel area threshold S th For comparison, if S o >S th , a wide-angle command will be sent, and the target pixel area in the market will be reduced. Conversely, the FPGA sends a telephoto command, and the target pixel area in the field of view will be increased.

[0039] As a preferred technical solution of the present invention, the monitoring device segments the target area based on the improved two-dimensional OTSU algorithm to obtain the pixel area where the target is located. o With threshold S th After comparison, it is converted into zoom steps and zoom direction.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] Dynamically optimize traffic flow and improve traffic efficiency

[0042] By combining real-time, multi-dimensional traffic data captured by monitoring equipment with a dynamic timing model driven by reinforcement learning, this system can accurately predict traffic flow trends and generate optimal signal strategies. Field tests have shown that, in complex intersection scenarios, average vehicle delays have been reduced by 30%-45%, and peak-hour capacity has increased by over 20%.

[0043] Central data coordination to improve overall traffic flow

[0044] By adjusting local data through Nash equilibrium, we can effectively overcome the defects of local data optimization but global suboptimality, further improve traffic efficiency and ensure stable operation around the clock.

[0045] Distributed edge computing architecture to reduce response delays

[0046] Relying on distributed traffic light controllers to achieve localized real-time decision-making, the delay in issuing signal control instructions is compressed to milliseconds, which is 5-8 times faster than the response speed of traditional centralized systems, meeting the immediate control needs of sudden traffic events (such as accidents and emergency vehicle priority passage).

[0047] Adaptive control of emergencies to improve road network resilience

[0048] Through regional signal coordination and linkage, the efficiency of suppressing congestion spread has been improved by 50%, and the overall traffic stability of the road network has been significantly enhanced.

[0049] Energy conservation, environmental protection and sustainable development

[0050] By reducing the frequency of vehicle starts and stops and idling time, the system can reduce fuel consumption of vehicles around intersections by 8%-12%, and simultaneously reduce carbon emissions, helping to promote the green transformation of urban transportation. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0052] Figure 1 It is a schematic diagram of the overall structure of the present invention;

[0053] Figure 2 It is a diagram of the cascaded expanded convolution structure used in the present invention;

[0054] Figure 3 is a flow chart of the method of the present invention;

[0055] In the figure: 1. Monitoring equipment; 2. Traffic light controller; 3. Central control system. DETAILED DESCRIPTION

[0056] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0057] Furthermore, if detailed description of known art is not necessary to illustrate the characteristics of the present invention, it will be omitted.

[0058] Example 1

[0059] like Figure 1-3As shown, the present invention provides a real-time monitoring and control system for traffic signals, comprising a monitoring device 1 arranged at each traffic intersection in an area, a signal light controller 2 wirelessly connected to the monitoring device 1, and a central control system 3 connected to all signal light controllers in the area, wherein the monitoring device 1 comprises a data acquisition module, a data preprocessing module, and a data feature extraction module, wherein the data acquisition module is used to capture the status of each intersection and send the captured image to the data preprocessing module, the data preprocessing module preprocesses the data and sends it to the data feature extraction module, the data feature extraction module extracts features from the data, extracts the status data of each intersection, and transmits the information to the signal light controller 2, the signal light controller 2 adopts a reinforcement learning algorithm to adjust the signal status of the traffic light, and sends the adjusted plan to the central control system 3, the central control system 3 collects the adjustment plans of each intersection, adopts Nash equilibrium to readjust the plan, and sends the adjusted plan to the signal light controller 2 again, and the signal light controller 2 adjusts the signal according to the adjustment plan issued by the central control system 3.

[0060] Specifically, signal adjustment is performed using the Q-reinforcement learning algorithm. This unsupervised learning algorithm uses rewards and penalties from the environment to autonomously discover strategies that maximize reward. First, monitoring equipment 1 collects status data from each intersection. This data includes vehicle saturation, traffic flow, and traffic light status.

[0061] The current state of the traffic lights at each intersection at time t and the saturation of the vehicles at the intersection are S t , the strategy adopted at time t is a t , Q(S t , a t ) is the total traffic volume at time t, α is the learning rate, which is adaptively adjusted based on the dynamic traffic load; γ is the discount rate, and the evaluation value of the strategy adopted at time t is r t , then the formula of the reinforcement learning algorithm of Q learning is:

[0062]

[0063] Reinforcement learning algorithms are used to obtain the local optimal strategy for each intersection, but each intersection only makes the traffic at that intersection smooth. However, the intersections influence and interact with each other, which leads to coordination problems between the intersections and even contradictions. Therefore, a central control system is needed to readjust them to achieve global smoothness. The decision of one intersection will inevitably be affected by the decisions of other intersections. Therefore, a game method is used to achieve equilibrium, so as to better coordinate the relationship between the intersections.

[0064] The calculation formula of Nash equilibrium is;

[0065]

[0066] Where, represents the strategy selected by the i-th signal lamp controller; represents the vector composed of the strategies of all signal lamp controllers except i; Q i represents the gain and loss of the subject, that is, the utility level, under the established strategy combination of the i-th signal lamp controller; S i represents the strategy space of the i-th signal lamp controller; s i represents a certain strategy of the i-th signal lamp controller, which is a subset of S i ; t represents the time.

[0067] In the monitoring device 1, feature extraction is performed by cascaded convolution for sampling to obtain a feature extraction map. Each time cascaded convolution cascades N extended convolutional layers, the size of the convolutional kernel is k×k, and the expansion rates are [r1,…,r i ,…,r n .

[0068] The expansion rate r~i~ of cascaded convolution satisfies r~i~ = 2^i^, so as to gradually expand the receptive field, extract multi-scale traffic features, adaptively capture the near-field vehicle density and far-field traffic flow trend through the dynamic expansion rate, and improve the feature extraction efficiency higher than that of the traditional CNN.

[0069] Define the maximum value of the expansion rate of the i-th layer in the extended convolutional kernel as R i , and the expansion rate r i ≤R i , and the calculation formula of R i is:

[0070]

[0071] And (R1 - 1) + (R2 - 1) < k, when R i > 0, R i ≠R i+1 ; the expansion rates r i within each group cannot form a common factor relationship; r i represents the expansion rate of the i-th layer in the extended convolutional kernel, and k is the size of the convolutional kernel.

[0072] Convolution is calculated using batch normalization, and then through the activation function Relu. The batch normalization calculation includes the following steps:

[0073] B1: Input the feature maps of the same batch into the batch normalization layer for calculation to obtain the mean value and variance

[0074] B2: for all x i , i∈1…m is standardized, and we get

[0075] B3: Yes Do a linear transformation and output the batch normalization result y i .

[0076] described The mean and variance The calculation formulas are The x i The standardized calculation formula is The y i The calculation formula is

[0077] Where m represents the batch capacity; ε represents a very small constant to avoid system errors caused by a denominator of 0; β and γ represent adjustment parameters.

[0078] The adjustment parameters β and γ are both trainable parameters in the neural network. The initial value of γ is 1, and the initial value of β is 0. The calculation formulas of γ and β are as follows: γ′←γ+Δγ; β′←β+Δβ;

[0079] Where y represents the model output image of this training round, represents the labeled image. η represents the learning rate, η = 0.0001.

[0080] The monitoring device sequentially performs grayscale processing, normalization, contrast-limited adaptive histogram equalization, and gamma nonlinear processing to obtain a preprocessed image.

[0081] The normalization steps include:

[0082] Auxiliary variables are screened using the 3σ principle. Subsequently, maximum and minimum normalization is performed to convert data of different dimensions into data of the same dimension. By mapping the data to [0, 1], the influence of different dimensions is eliminated, making different indicators comparable and improving the convergence speed of the model.

[0083]

[0084] in x t max 、x t minThey represent the normalized time series, the maximum value in the data, and the minimum value in the data, respectively.

[0085] The normalization comprises the following steps:

[0086] Let X = {x 1 , x 2 ,...,x n} is an image dataset; the Z score normalization method is used to normalize the dimension of the dataset X; the Z score normalization formula is as follows:

[0087] Xnorm=(X-μ) / σ

[0088] Where μ is the mean of x and σ is the standard deviation of x. In this case, x has both positive and negative values, and the mean of x is 0 and the standard deviation is 1.

[0089] Then, each image data in the dataset X is normalized by min-max, and the value of x is remapped to the range of 0 to 255. The min-max normalization formula is as follows:

[0090]

[0091] Where x i ∈X norm , i∈[1,2,...,n].

[0092] Contrast-limited adaptive histogram equalization uses the CLAHE algorithm to enhance the feature and background contrast of the entire data set, and then uses gamma nonlinearization to perform nonlinear operations on the luminance or tristimulus values ​​in the image to obtain the preprocessed image.

[0093] Normalization and contrast-limited adaptive histogram equalization and gamma nonlinear processing are used to make the processed image clearer, improve the resolution, and make the subsequent feature extraction more accurate.

[0094] After the monitoring equipment obtains the regional image containing features, it uses the improved two-dimensional OTSU algorithm to segment the regional image, obtain the target area and the background area respectively, and calculate the pixel area S of the target area. o and with the standard pixel area threshold S th For comparison, if S o >S th , a wide-angle command will be sent, and the target pixel area in the market will be reduced. Conversely, the FPGA sends a telephoto command, and the target pixel area in the field of view will be increased.

[0095] The monitoring device segments the target area based on the improved two-dimensional OTSU algorithm to obtain the pixel area where the target is located, S o With threshold S thAfter comparison, it is converted into zoom steps and zoom direction.

[0096] Focusing the monitoring equipment can further improve image clarity, make subsequent data processing more accurate, and reduce errors.

[0097] The data used in the experiment comes from the traffic flow collected by the detector throughout the day, with an interval of 5 minutes. Figure 1 As shown in the figure, Intersections 1, 2, 3, and 4 are four adjacent intersections. The westbound intersection of Intersection 1 was selected as the traffic flow monitoring point, and the average delay time after vehicles passed through both intersections was used as the evaluation metric. To verify the effectiveness of the algorithm under different traffic conditions, this paper used traffic flow data from different time periods, with 7:30-8:30 as the peak period and 9:30-10:30 as the low-peak period. The road section length s = 1500 meters at each of the four intersections was used for comparison. The results were run over approximately 1000 cycles, considering only the motor vehicle lanes. All intersections were set up with four phases, with the proportions of left-turning, through-going, and right-turning vehicles being 30%, 40%, and 30% respectively. Coordinated control is performed during off-peak and peak periods, and then compared with the delay time using timing control. The simulation results are the average delay time of vehicles passing through the intersection. The delay times during off-peak periods are 5.2s and 22.9s, respectively, and the delay times during peak periods are 30.9s and 47.9s, respectively. From the results, it can be seen that compared with ordinary timing control, the optimization efficiency of this application is above 77%.

[0098] The present invention uses monitoring equipment to capture multi-dimensional traffic data in real time. Combined with a dynamic timing model driven by reinforcement learning, it can accurately predict traffic flow trends and generate optimal signal strategies. Local data can be adjusted through Nash equilibrium, effectively overcoming the defect of local data optimization but global suboptimality, further improving traffic efficiency and ensuring stable operation around the clock. Through regional signal coordination and linkage, the efficiency of suppressing congestion spread can be increased by 50%, and the global traffic stability of the road network can be significantly enhanced.

[0099] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A traffic signal real-time monitoring and control system, characterized in that: The invention comprises a monitoring device (1) arranged at each traffic intersection in an area, a signal light controller (2) connected to the monitoring device (1) by wireless communication, and a central control system (3) connected to all signal light controllers in the area, wherein the monitoring device (1) comprises a data acquisition module, a data preprocessing module, and a data feature extraction module, wherein the data acquisition module is used to capture the status of each intersection and send the captured image to the data preprocessing module, wherein the data preprocessing module preprocesses the data and sends the image to the data feature extraction module, wherein the data feature extraction module extracts the feature of the data, extracts the status data of each intersection, and transmits the information to the signal light controller (2), wherein the signal light controller (2) adopts a reinforcement learning algorithm to adjust the signal status of the traffic light and sends the adjusted scheme to the central control system (3), wherein the central control system (3) collects the adjustment schemes of each intersection and readjusts the schemes by adopting Nash equilibrium, and sends the adjusted schemes to the signal light controller (2) again, wherein the signal light controller (2) performs signal adjustment according to the adjustment scheme sent by the central control system (3).

2. A traffic signal real-time monitoring and control system according to claim 1, characterized in that: The status data of each intersection includes the saturation of vehicles at each intersection, flow rate, and status of traffic lights.

3. A traffic signal real-time monitoring and control system according to claim 2, characterized in that: The signal light controller (2) comprises a data collating unit and a calculation and analysis unit; The data sorting unit sorts the feature data extracted by the data feature extraction module, divides the time into several moments according to the strategy adopted at each moment, obtains the current state of the traffic lights at each intersection at time t, the saturation of vehicles at the intersection, and the total flow at time t+1, and sends the data to the calculation and analysis unit; the strategy is the action instruction of the traffic light, and the action instruction includes two states: the state of the traffic light is switched (such as the traffic light is switched to green) and the state is unchanged. The calculation and analysis unit uses the current state of the traffic lights at each intersection at time t and the saturation data of the vehicles at the intersection as the state data S of the intersection t , a t The strategy adopted at time t is the sum of the traffic at each time t as the reward value Q(S t , a t ), and set the evaluation value of the strategy taken at time t to r t+1 , when the total flow at time t+1 is less than the total flow at time t, the evaluation value r t+1 Greater than 0, then state S t a t Should be re-encouraged, otherwise weakened. Put the above data into the reinforcement learning model for training, then the formula of the reinforcement learning algorithm of Q learning is: The learning rate is a constant between 0 and 1, γ is the discount rate, and γ is equal at each moment; S t+1 is the current status of the traffic lights at each intersection at time t+1 and the saturation of vehicles at the intersection.

4. A traffic signal real-time monitoring and control system according to claim 1, characterized in that: The central control system (3) includes a data aggregation unit and a game decision unit; The data aggregation unit aggregates the optimal calculation results obtained by the traffic light controllers (2) at each intersection, marks the traffic light controllers at each intersection, records them as i, takes the set of all possible strategies or actions of the traffic light controllers at each intersection as S, takes the traffic flow situation under each given strategy combination as Q, and sends the data to the game decision unit; The game decision unit uses the traffic light controllers at each intersection to play a game to obtain the global optimal solution. (The reinforcement learning algorithm is used to obtain the local optimal strategy for each intersection, but each intersection only ensures smooth traffic at that intersection. However, each intersection influences and interacts with each other, which leads to coordination problems and even contradictions between the intersections. Therefore, a central control system is needed to readjust them to achieve global smooth traffic. The decision of one intersection is inevitably affected by the decisions of other intersections. Therefore, a game method is used to achieve equilibrium, thereby better coordinating the relationship between the intersections). The calculation formula of the game is: in, represents the strategy selected by the i-th traffic light controller; represents the vector of strategies of all traffic light controllers except i; Q i represents the utility level of the i-th traffic light controller; S i represents the strategy space of the i-th traffic light controller; s i Represents a strategy of the i-th traffic light controller, which is S i A subset of ; t represents the time.

5. A traffic signal real-time monitoring and control system according to claim 1, characterized in that: The feature extraction in the monitoring device (1) adopts cascade convolution to perform sampling to obtain a feature extraction graph. The cascade convolution cascades N expansion convolution layers each time, the size of the convolution kernel is k×k, and the expansion rates are [r1,…,r i ,…,r n ].

6. A traffic signal real-time monitoring and control system according to claim 1, characterized in that: The expansion rate r~i~ of the cascade convolution satisfies r~i~=2^i^, so as to gradually expand the receptive field and extract multi-scale traffic features.

7. A traffic signal real-time monitoring and control system according to claim 1, characterized in that: The monitoring device sequentially performs grayscale processing, normalization, contrast-limited adaptive histogram equalization, and gamma nonlinear processing to obtain a preprocessed image, wherein the normalization step includes: Auxiliary variables are screened using the 3σ principle. Subsequently, maximum and minimum normalization is performed to convert data of different dimensions into data of the same dimension. By mapping the data to [0, 1], the influence of different dimensions is eliminated, making different indicators comparable and improving the convergence speed of the model. in x t max 、x t min They represent the normalized time series, the maximum value in the data, and the minimum value in the data, respectively.

8. A traffic signal real-time monitoring and control system according to claim 1, characterized in that: The contrast-limited adaptive histogram equalization adopts the CLAHE algorithm to enhance the feature and background contrast of the entire data set, and then performs nonlinear operations on the luminance or tristimulus values ​​in the image by gamma nonlinearization to obtain a preprocessed image.

9. A traffic signal real-time monitoring and control system according to claim 1, characterized in that: After the monitoring device obtains the regional image containing the features, the improved two-dimensional OTSU algorithm is used to segment the regional image to obtain the target area and the background area respectively, and the pixel area S of the target area is calculated. o and with the standard pixel area threshold S th For comparison, if S o >S th , a wide-angle command will be sent, and the target pixel area in the market will be reduced. Conversely, the FPGA sends a telephoto command, and the target pixel area in the field of view will be increased.

10. A traffic signal real-time monitoring and control system according to claim 1, characterized in that: The monitoring device segments the target area based on the improved two-dimensional OTSU algorithm to obtain the pixel area where the target is located, S o With threshold S th After comparison, it is converted into zoom steps and zoom direction.

Images