A smart energy-saving lamp pole group control method based on recurrent neural network algorithm

Abstract

The present invention discloses a control method for a smart energy-saving lamp pole group based on a recurrent neural network algorithm, the method comprising: collecting real-time road information through a video acquisition camera, and obtaining a preprocessed image through block matching; using a trained improved recurrent neural network system to identify video image information; setting a virtual section to calculate the instantaneous speed of a vehicle identified in the video image; controlling the smart lamp pole in real time based on the obtained data; turning on the lighting when the ambient illumination is lower than 20lx for 10 minutes, otherwise turning it off; turning on a loudspeaker and sounding an alarm when it is determined according to the data that the vehicle has dangerous driving behaviors such as speeding or fatigue driving, and converting the lighting modes of the lighting lamps in two lamp pole groups in the direction of traffic flow into a strobe mode for 5 seconds; when it is determined that a traffic incident occurs on the road section, all LED display screens in the opposite direction of the traffic flow of the line will publish the road congestion information ahead in real time to prevent the rear vehicles from avoiding in time and further deteriorating the traffic situation.

Description

Technical Field

[0001] The present invention relates to the technical field of lamp pole control, and in particular to a smart energy-saving lamp pole group control method based on a recurrent neural network algorithm. Background Art

[0002] A lamp pole is a lighting facility used on roads. Most existing lamp poles are non-intelligent lamp poles with fewer functions, especially in terms of lighting on expressways. Vehicles on expressways travel at high speeds and have high transportation efficiency. If a traffic incident occurs, it is easy to cause a vicious cycle of expressway congestion if the vehicles behind do not receive timely road information. At the same time, in order to achieve the purpose of green environmental protection, energy conservation and emission reduction, the present invention uses video recognition technology to identify and calculate parameters such as ambient illumination, vehicle passing, and vehicle speed, so that lighting facilities can be adjusted in real time to meet the needs.

[0003] At the same time, when a traffic incident is identified on a certain road section through video recognition technology, the information release device is used to release real-time road information, so that vehicles traveling on the expressway can easily and quickly obtain key information, alleviating traffic pressure. Summary of the invention

[0004] In order to solve the above problems and overcome the shortcomings of the prior art, the present invention provides a smart energy-saving lamp pole group control method based on a recurrent neural network algorithm.

[0005] To achieve the above object, the present invention is implemented by the following technical solutions:

[0006] A smart energy-saving lamp pole group control method based on a recurrent neural network algorithm comprises the following steps:

[0007] S1. Capture continuous multi-frame video image information through a video acquisition camera, segment the video image information using a block matching method, and obtain pre-processed image information.

[0008] S2. The segmented pre-processed image information is transmitted to the recurrent neural network (RNN) algorithm module, and the vehicle in the video image is identified and tracked through the trained recurrent neural network system.

[0009] S3. Track the motion trajectory of the identified vehicle, establish motion connections between multiple consecutive frames of images, and calculate the speed of the vehicle identified in the image.

[0010] The specific method is to use the headlights of the vehicle identified in the video as the base point, and calculate the average speed of the vehicle in this section based on the time it takes for it to pass through two specified virtual sections, which is equivalent to the instantaneous speed of the vehicle at this point.

[0011] S4. When the video acquisition camera of lamp pole A in the reverse traffic flow direction detects that a vehicle passes by, lamp pole A is used as the starting point to convert the brightness of the lamps of the four lamp pole groups in front of the traffic flow into lighting mode. When lamp pole B of the next lamp pole group under lamp pole A detects that a vehicle passes by the position of the lamp pole, the lamp pole group where lamp pole A is located is converted into energy-saving mode and the illumination is reduced. The priority of lighting mode is higher than that of energy-saving mode.

[0012] S5. By calculating the vehicle speed in the video image, determine whether the driver has dangerous driving behavior, including speeding and fatigue driving. If it is detected that the instantaneous speed of the vehicle exceeds 20% of the maximum speed limit of the section or the vehicle speed exceeds the speed limit and the standard deviation of the speed in 20 consecutive groups of lamp poles is less than 2.5km / h, it is considered as fatigue driving, the speaker is turned on to sound an alarm, and the lighting mode of the lights in the two lamp pole groups in the forward direction of the traffic flow is converted to strobe mode for 5 seconds. If the instantaneous speed of the vehicle is detected to be less than 10km / h, it is considered that an emergency has occurred at this location, and the lights of the two lamp pole groups in the opposite direction of the traffic flow are converted to gentle flashing.

[0013] The stroboscopic mode is a rapid switch between the energy-saving mode and the lighting mode at an interval of 0.25 seconds; the gentle flashing mode is a gentle switch between the energy-saving mode and the lighting mode at an interval of 0.75 seconds; wherein the illumination of the lighting mode is 35lx, and the illumination of the energy-saving mode is 20lx.

[0014] S6. By detecting the illumination of the video image information, if the illumination of the collected video information remains ≤20lx for more than 10 minutes, the lighting is turned on; if the illumination of the collected video information remains ≥20lx for more than 10 minutes, the lighting is turned off.

[0015] S7. By acquiring map information, the expressway ramp and route destination are published on the LED display screen. By detecting traffic accidents and road congestion, if a traffic accident is detected on a road section, the LED display screens of all lamp pole groups in the opposite direction of the traffic flow of the line will publish the road congestion information ahead in real time. Through guiding display, vehicles are guided to choose the appropriate ramp to leave the expressway.

[0016] The method for determining traffic accidents and road congestion is to detect the vehicle speed. If the instantaneous speed of the vehicle is detected to be less than 10km / h, it is determined that an emergency has occurred at this location. At this time, the corresponding guidance image is displayed, otherwise ordinary road information is displayed.

[0017] Furthermore, the lamp pole group is specifically: assuming that the lamp pole spacing within the lamp pole group is l, the group spacing between adjacent lamp pole groups is l, and the number of lamp poles in the lamp pole group is n, then the number of lamp poles included in a straight section with a length of L is N=L / l+1, and the number of lamp pole groups G=L / nl, wherein the lamp pole spacing within the lamp pole group and the lamp pole group spacing l are 35m, and the number of lamp poles in the lamp pole group is n, which is 2.

[0018] Furthermore, the criteria for determining dangerous driving behaviors such as speeding combine relevant standards in the Road Traffic Safety Law of the People's Republic of China and specific road conditions to determine that a vehicle speed reaching 120% of the speed limit can be identified as dangerous driving behaviors such as speeding.

[0019] Furthermore, the criteria for judging dangerous driving behaviors such as fatigue driving are determined through simulation experiments. When the vehicle speed exceeds the speed limit and the standard deviation of the vehicle speed within 20 consecutive groups of lamp poles is less than 2.5 km / h, it can be identified as dangerous driving behaviors such as fatigue driving.

[0020] Furthermore, the criteria for determining whether the road section needs lighting are based on the "Technical Conditions for Highway Lighting" and specific circumstances to determine whether the road section needs lighting, namely whether the illumination of the road section is lower than 201x, and the illumination of the lamp in the lamp pole lighting mode is determined to be 351x.

[0021] Furthermore, in order to increase the transmission and processing speed of information and improve the real-time performance of the system, a 5G module is used to utilize its large bandwidth, low latency and other characteristics to improve the transmission speed. A recurrent neural network algorithm is used in combination with a long short-term memory neural network (LSTM) algorithm for optimization. At the cost of increasing storage space, high processing speed and accuracy are obtained, further improving the real-time performance of the system. At the same time, since the target identification is simple and the space occupied by the extracted eigenvalues ​​is small, the cost of the additional storage space can be ignored.

[0022] Furthermore, the RNN algorithm module optimized by the long short-term memory neural network in step S2 obtains the output item of the time node by combining the image data of the previous time node with the image data of the current time node. The RNN-LSTM algorithm introduces a loop structure of the hidden layer, which makes the algorithm memorable. The previous time node can affect the result of the subsequent node, and can effectively shield the influence of accidental situations on the output, such as occlusion, etc., which is very suitable for processing continuous video image data;

[0023] Furthermore, the RNN-LSTM algorithm module specifically includes the following steps:

[0024] After receiving the segmented preprocessed image information, the S31.RNN-LSTM algorithm module inputs the sample image into the trained neural network system as the input item x of the network structure. If in the time series 1-T, the sample image of this time node is the input item x of the input layer. t At the t time series node, the hidden state h corresponding to the hidden layer t The hidden state h of the previous node can be t-1 , the input item x of the current node t 、Long-term memory knowledge element C t-1 , and the weight matrices U, W, and X are calculated, that is, the weight matrix Ux is applied to the input item, the long-term memory knowledge element, and the hidden state of the previous node: t +WC t-1 +Xh t-1 , get the short-term memory forgetting element S t , and then apply the weight matrix Ux to the input, short-term memory forget element, and the previous node hidden state: t +WS t +Xh t-1 , get long-term memory knowledge element C t And store it, then apply the weight matrix Ux to the input item, long-term memory knowledge element and the hidden state of the previous node: t +WC t +Xh t-1 , and use the activation function to activate the hidden state h t ; Similarly, the output item of this node is composed of the hidden state h of the node t And the weight matrix V and activation function g are calculated:

[0025] o t =g(Vh t )

[0026] From this, we can get the predicted output value o of each sequence node t , combined with the loss function E t Get the actual output y of the sequence node t Determine whether it is a vehicle passing by and track the identified vehicle.

[0027] S32. Calculate the loss term of each node through the back propagation method. Suppose the loss term of the mth layer at time t is It is propagated along the layer and timeline to the previous layer and previous moment of the input layer and hidden layer, and the current layer of the long-term memory knowledge element, and we get and Assume n t is the weighted input of the system at the t sequence node, that is:

[0028] n t =Uxt +WC t +Xh t-1

[0029] h t =n t

[0030] By The term is decomposed into Calculate the loss term α and use the back propagation rule to get the loss term at any time k

[0031]

[0032] The function of diag[x] is to create a diagonal matrix based on the x vector.

[0033] S33. Calculate the gradient of the loss function E with respect to each weight matrix U, W, X, and V.

[0034] The gradient of the loss function E with respect to the weight matrix U is calculated by substituting the loss term at time t into With input x t Multiply, repeat the operation over the time series and sum the gradient of the obtained weight matrix U

[0035] Similarly, the gradient of the loss function E with respect to the weight matrix W is calculated by taking the loss term at time t and the long-term memory knowledge element C at this moment t Multiply, repeat the operation over the time series and sum the gradient of the obtained weight matrix W Similarly, the gradient of the loss function E with respect to the weight matrix X is calculated by taking the loss term at time t and the hidden item h at the previous moment t-1 Multiply, repeat the operation over the time series and sum the gradient of the obtained weight matrix X

[0036] The gradient of the loss function E with respect to the weight matrix V is:

[0037]

[0038] The gradient formulas are all summations, which may cause gradient vanishing or gradient exploding problems due to the length of the sequence.

[0039] Furthermore, the RNN-LSTM algorithm in step S23 may have the problem of gradient vanishing or gradient exploding due to the length of data information. This method uses a GRU network structure to improve the RNN-LSTM, specifically including the following steps:

[0040] S431. Based on the original RNN-LSTM, GRU retains the dual-input and dual-output structure, adds two internal gate structures, "update gate" and "reset gate", and obtains the hidden state h through the previous sequence node transmission. t-1 and the input item x of the current node t And the weight matrix Q, activation function β (control N, S values ​​belong to the range of (0, 1)) to obtain the gating signals N, S, where N controls the update gate (renewgate) and S controls the reset gate (resetgate).

[0041]

[0042]

[0043] S432. Reset the data by obtaining the “reset gate” gating signal S:

[0044] h′ t-1 =h t-1 *S

[0045] The reset hidden state h′ t-1 The input item x of the current node t Combined with the function th, the value is controlled in the interval (-1, 1) to obtain h′ t :

[0046]

[0047] Using the obtained “update gate” gating signal N and h′ t h for this node t Perform "update" and use the value of N to control the influence of the previous node on this node (the value of N is in the range of (0, 1), and the closer the value is to 0, the smaller the influence):

[0048] h t =N(h t-1 -h′ t )+h′ t

[0049] Furthermore, according to the intelligent energy-saving lamp pole group control method based on the recurrent neural network algorithm of claim 3, it is characterized in that the RNN-LSTM algorithm in step S31 can reduce the computational complexity and improve the computational efficiency through time parallel computing. The specific steps are as follows:

[0050] S51. Hidden state h of the previous node t-1 and the input item x of the current node t Apply weight matrix: Ux t +Xh t-1, directly solve in parallel to obtain the intermediate variable i of the forget gate t , input gate intermediate variable f t and z t , output gate intermediate variable o t ;

[0051] S52. Forget gate intermediate variable i obtained by parallel calculation t and long-term memory knowledge element C t-1 Apply weight matrix: i t +WC t-1 , get the short-term memory forgetting element S t ; At the same time, the input gate intermediate variable f t and z t Added together, it is the intermediate variable c′ of the long-term memory knowledge element t ;

[0052] S53. Forgetting elements of short-term memory t and the intermediate variable c′ of the long-term memory knowledge element t Apply weight matrix: c′ t +WS t , get long-term memory knowledge element C t And store it, then output the intermediate variable o t 、Long-term memory knowledge element C t Apply the weight matrix: o t +WC t , and use the activation function to activate the hidden state h t ; The subsequent activation function steps are the same as in step S31.

[0053] Compared with the prior art, the present invention provides a smart energy-saving lamp pole group control method based on a recurrent neural network algorithm, which has the following beneficial effects: (1) By intelligently controlling the on and off, illumination, and lighting mode of the lighting device, the lighting function is fully guaranteed while taking into account energy conservation and emission reduction, low carbon and environmental protection, and the purpose of warning the rear vehicles by controlling the lighting mode when a special traffic incident occurs. (2) By publishing real-time traffic information through the LED display screen, vehicles traveling on the expressway can accurately obtain the road information and traffic status ahead, obtain the specific information of the traffic incident ahead in a timely manner, and correctly guide the vehicle, thereby reducing the burden of event processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a flow chart of a method for controlling a smart energy-saving lamp pole group based on video recognition technology;

[0055] Figure 2 A logic diagram of the recurrent neural network of step S2 of the present invention;

[0056] Figure 3 A flow chart of energy-saving lighting control according to the present invention;

[0057] Figure 4 It is a schematic diagram of the light pole layout and light pole group settings;

[0058] Figure 5 This is a schematic diagram of the present invention detecting vehicle passing and dangerous driving behavior corresponding mode conversion when the span of the lamp pole group is 50m;

[0059] Figure 6 This is a schematic diagram of the corresponding mode conversion when the span of the lamp pole group is 50m when detecting an emergency event according to the present invention. DETAILED DESCRIPTION

[0060] In order to make the technical means, creative features, objectives and effects of the present invention easy to understand, the following is a brief introduction to the present invention. Figure 1-3 Specific embodiments are described and illustrated.

[0061] Embodiment 1:

[0062] The control principle of the invention specifically includes the following steps:

[0063] Step 1: Collect continuous multi-frame video image information through the video acquisition camera, perform block matching segmentation on the video image, obtain pre-processed image information and transmit it to the recurrent neural network (RNN) algorithm module.

[0064] Step 2: Recognize the pre-processed image after segmentation. Specifically, the RNN-LSTM algorithm module receives the pre-processed image information after segmentation and inputs the sample into the trained neural network system as the input item x of the network structure. The t time series node is the input item x of the input layer. t 、Long-term memory knowledge element C t-1 , combined with the matrices U, W, and X, the hidden state h of the hidden layer of the previous sequence node t-1 Calculate the hidden state h corresponding to the hidden layer t , the hidden state h of the node t And the matrix V can be calculated to get the output item o of this node t , and then combine the loss function E to calculate the true output y of the node t , where h t The GRU structure algorithm module has been "updated" and "reset".

[0065] Step three, track the movement trajectory of the vehicle identified in the video with its headlights as the base point, establish motion connections between multiple consecutive frames of images, and calculate the average speed of the vehicle in this section based on the time it takes for it to pass through two specified virtual sections, which is equivalent to the instantaneous speed of the vehicle at this point.

[0066] Step 4: Control the lighting changes of lighting devices to achieve energy conservation, emission reduction and traffic assistance, specifically:

[0067] By detecting the illumination of the video image information, if the illumination of the collected video information remains ≤20lx for more than 10 minutes, the lighting will be turned on; if the illumination of the collected video information remains ≥20lx for more than 10 minutes, the lighting will be turned off;

[0068] When the video acquisition camera on lamp pole A in the reverse traffic flow direction detects a vehicle passing by, the lamp pole A is used as the starting point to convert the brightness of the lamps of the four lamp pole groups in front of the traffic flow into the lighting mode. When lamp pole B in the next lamp pole group under lamp pole A detects that a vehicle has passed the lamp pole position, the lamp pole group where lamp pole A is located is converted into the energy-saving mode and the illumination is reduced. The priority of the lighting mode is higher than the energy-saving mode. By calculating the speed of the vehicle in the video image, it is determined whether the driver has dangerous driving behavior, including speeding and fatigue driving. If it is detected that the instantaneous speed of the vehicle exceeds the maximum speed limit of the section by 20%, it is considered as speeding. If the vehicle speed exceeds the speed limit and the standard deviation of the speed in 20 consecutive lamp pole groups is less than 2.5km / h, the driver's actions are obviously too few and the speed is high, it is considered as fatigue driving, and the speaker is turned on to sound an alarm.

[0069] The lighting mode of the lights in the two lamp pole groups in front of the traffic flow will be converted to strobe mode for 5 seconds. If the instantaneous speed of the vehicle is detected to be less than 10km / h, it is determined that an emergency has occurred at this location, and the lights of the two lamp pole groups in the opposite direction of the traffic flow will be converted to gentle flashing to warn the vehicles behind to slow down in time.

[0070] Step five, by obtaining map information, the expressway ramp and route destination are published on the LED display screen. By detecting traffic accidents and road congestion, if a traffic accident is detected on a road section, all LED display screens in the opposite direction of the traffic flow will publish the road congestion information ahead in real time. Through guiding display, vehicles are guided to choose a suitable ramp to leave the expressway. Specifically: when a certain road section ahead is identified as having a traffic incident, the LED display screen will publish real-time warning information about the traffic incident on the road section ahead and the server, ramp and destination leading to the ramp in between, so that the driver can choose a suitable route.

[0071] Example 2:

[0072] Reference Figure 3-6 , an energy-saving control lighting method of a smart energy-saving lamp pole group based on a recurrent neural network algorithm specifically includes the following steps:

[0073] Z1. Collect real-time video image information through the video acquisition camera, and identify the ambient illumination through the video image information. When the ambient illumination remains ≤20lx for more than 10 minutes, it indicates that the ambient environment needs to be illuminated, and the lights are turned on at this time; when the ambient illumination remains ≥20lx for more than 10 minutes, it indicates that the ambient environment no longer needs to be illuminated, and the lights are turned off at this time. The real-time changed lighting time matches the ambient brightness, fully ensuring that the lighting needs are met while achieving the purpose of energy saving and environmental protection;

[0074] Z2. When the lighting is turned on, the lighting mode defaults to the energy-saving mode. The illumination in the energy-saving mode only meets the minimum limit of the lighting function. When the lighting is turned on, the video processing module is used to identify the vehicle. When a vehicle is identified passing by, the lighting lamps in the four lamp pole groups along the traffic flow direction are converted to the lighting mode to prevent the lighting from suddenly becoming brighter and causing adverse effects. The process of converting from the energy-saving mode to the lighting mode is a gradual brightening process, and it is not necessary to convert from the lighting mode to the energy-saving mode; the specific steps of the mode conversion control are:

[0075] When the video acquisition camera of lamp pole A in the reverse traffic flow direction detects that a vehicle has passed, the lamp pole A is used as the starting point, and the brightness of the lamps of the lamp poles of the four lamp pole groups in front of the traffic flow is converted to the lighting mode. When the lamp pole B of the lamp pole group under lamp pole A detects that a vehicle has passed the position of the lamp pole, the lamp pole group where lamp pole A is located is converted to the energy-saving mode, and the illumination is reduced. The priority of the lighting mode is higher than the energy-saving mode, that is, the purpose of energy saving is achieved when the lighting conditions are guaranteed;

[0076] Z3. Calculate the vehicle speed through video detection, set the speed threshold M = speed limit × 120%, when the detected vehicle speed exceeds the speed threshold M, that is, the vehicle speed exceeds the speed limit by 20%, the vehicle is identified as speeding dangerous driving behavior, and the lighting mode of the two lamp poles in the traffic flow direction is converted to strobe mode for 3 seconds;

[0077] Z4. Calculate the vehicle speed through video detection, set the speed threshold N = speed limit, when the detected vehicle speed exceeds the speed threshold N and is lower than the threshold M, that is, exceeds the speed limit but exceeds the speed limit by less than 20% and the standard deviation of the vehicle speed in 20 consecutive groups of light poles is less than 2.5km / h, the vehicle is identified as a dangerous driving behavior such as fatigue driving, and the lighting mode of the light poles in the two light pole groups along the traffic flow direction is converted to strobe mode for 3 seconds;

[0078] That is, when light pole C detects dangerous driving behavior of a vehicle, the lighting mode of the two light pole groups along the traffic flow direction is converted to strobe mode for 3 seconds;

[0079] Z5. Calculate the vehicle speed through video detection and set the speed threshold L = 10km / h. When the vehicle speed detected by the light pole E is lower than the speed threshold L, that is, the vehicle speed is lower than 10km / h, it is considered that an emergency has occurred at this location, and the lights in the two light pole groups in the opposite direction of the traffic flow, that is, the FG section lights, are converted into moderate flashing to warn the rear vehicles to slow down in time;

[0080] Z6. The stroboscopic mode is a rapid switch between the energy-saving mode and the lighting mode at an interval of 0.25 seconds; the gentle flashing mode is a gentle switch between the energy-saving mode and the lighting mode at an interval of 0.75 seconds.

[0081] Example 3:

[0082] The above implementation plans and examples are all based on expressways. If this control method is used in mountainous areas, some corresponding changes need to be made. This plan is changed based on the characteristics of mountainous areas such as small traffic volume, high risk, and complex terrain. Figure 4 The details are as follows:

[0083] 1. When controlling the light to switch between energy-saving mode and lighting mode, the light poles at the bends belong to the bend light pole group. The pole spacing in the light pole group is 1 / 2 of that in the straight section, and the light poles in the light pole group always remain in the lighting mode after the lighting is turned on. The purpose of this change is to maintain good lighting conditions at dangerous nodes on the road section and reduce the danger of the road;

[0084] 2. As the danger level of mountainous roads is relatively high, the road section’s tolerance for speed should be appropriately reduced, which is reflected in the change of the plan, that is, setting the speed threshold M s = speed limit × 110%. When the detected vehicle speed exceeds the speed threshold M, that is, the vehicle speed exceeds the speed limit by 10%, the vehicle is deemed to be in dangerous driving behavior of speeding, and the lighting mode of the lamp poles in the front two lamp pole groups is converted to strobe mode for 3 seconds;

[0085] Set speed threshold N = speed limit. When the detected vehicle speed exceeds speed threshold N and is lower than threshold M, that is, exceeds speed limit but exceeds speed limit by less than 10% and the standard deviation of vehicle speed in 15 consecutive groups of light poles is less than 2.5 km / h, the vehicle is identified as fatigue driving-related dangerous driving behavior, and the lighting mode of the light poles in the front 2 light pole groups is converted to strobe mode for 3 seconds;

[0086] 3. The LED display screen often displays key information such as parking locations within the road section, dangerous sections where accidents often occur, etc.

[0087] The above is only an embodiment of the present invention. The common sense such as the known specific structure and features in the scheme is not described in detail here. The ordinary technicians in the relevant field know all the common technical knowledge in the field to which the invention belongs before the application date or the priority date, can know all the existing technologies in the technical field, and have the ability to apply the conventional experimental means before that date. The ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known methods should not become obstacles for the ordinary technicians in the relevant field to implement this application. It should be pointed out that for the technicians in this field, several deformations and improvements can be made without departing from the structure of the present invention, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The protection scope required by this application shall be based on the content of the claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A smart energy-saving lamp pole group control method based on a recurrent neural network algorithm, characterized in that: The following steps are involved: S1. Collecting multiple frames of continuous video image information through a video acquisition camera, preprocessing the video image information to obtain preprocessed image information; S2. The pre-processed image information is transmitted to the recurrent neural network (RNN) algorithm module, and the vehicle in the video image is identified and tracked through the trained recurrent neural network system; S3. Track the motion trajectory of the identified vehicle, establish motion connections between consecutive multiple frames of images, and calculate the speed of the vehicle identified in the image; S4. When the video acquisition camera of lamp pole A in the reverse traffic flow direction detects that a vehicle has passed, the lamp pole A is used as the starting point, and the brightness of the lamps of the lamp poles of the four lamp pole groups in front of the traffic flow are converted to the lighting mode. When the lamp pole B of the lamp pole group under lamp pole A detects that a vehicle has passed the position of lamp pole B, the lamp pole group where lamp pole A is located is converted to the energy-saving mode, and the illumination is reduced. The priority of the lighting mode is higher than the energy-saving mode. S5. By calculating the vehicle speed in the video image, determine whether the driver has dangerous driving behavior, including speeding and fatigue driving. If the vehicle's instantaneous speed is detected to exceed 20% of the maximum speed limit of the road section or the vehicle's speed exceeds the speed limit and the standard deviation of the speed in 20 consecutive groups of light poles is less than 2.5km / h, it is considered as fatigue driving, the speaker is turned on to sound an alarm, and the lighting mode of the lights in the two light pole groups in the forward direction of the traffic flow is converted to strobe mode for 5 seconds. If the vehicle's instantaneous speed is detected to be less than 10km / h, it is considered that an emergency has occurred at this location, and the lights of the two light pole groups in the reverse direction of the traffic flow are converted to moderate flashing; S6. By detecting the illumination of the video image information, if the illumination of the collected video information remains ≤20lx for more than 10 minutes, the lighting will be turned on; if the illumination of the collected video information remains ≥20lx for more than 10 minutes, the lighting will be turned off; S7. By acquiring map information, the expressway ramp and route destination information are published on the LED display screen. By detecting traffic accidents and road congestion, if a traffic accident is detected on the road section, the LED display screens of all the lamp pole groups behind the road section will publish the road congestion information ahead in real time, and guide vehicles to choose the appropriate ramp to leave through guiding display; Specifically, the smart energy-saving lamp pole group is as follows: assuming that the distance between lamp poles in a lamp pole group is l, the distance between adjacent lamp pole groups is l, and the number of lamp poles in a lamp pole group is n, then the number of lamp poles included in a straight section with a length of L is N=L / l+1, and the number of lamp pole groups is G=L / nl; In order to increase the transmission and processing speed of information and improve the real-time performance of system operation, a 5G module is used to improve the transmission speed by taking advantage of its large bandwidth and low latency. The recurrent neural network (RNN) algorithm is used in combination with the long short-term memory neural network (LSTM) algorithm for optimization. The RNN-LSTM algorithm module in step S2 specifically includes the following steps: After receiving the segmented preprocessed image information, the RNN-LSTM algorithm module inputs the sample image into the trained neural network system as the input item x of the network structure. At the time node lT, the sample image at this time node is the input item x of the input layer. t ; At the t time series node, the hidden state h corresponding to the hidden layer t The hidden state h of the previous node can be t-1 , the input item x of the current node t 、Long-term memory knowledge element C t-1 , and the weight matrices U, W, and X are calculated, that is, the weight matrix Ux is applied to the input item, the long-term memory knowledge element, and the hidden state of the previous node: t +WC t-1 +Xh t-1 , get the short-term memory forgetting element S t , and then apply the weight matrix Ux to the input, short-term memory forget element, and the previous node hidden state: t +WS t +Xh t-1 , get long-term memory knowledge element C t And store it, then apply the weight matrix Ux to the input item, long-term memory knowledge element and the hidden state of the previous node: t +WC t +Xh t-1 , and use the activation function to activate the hidden state h t ; Similarly, the output item of this node is composed of the hidden state h of the node t And the weight matrix V and activation function g are calculated: o t =g(Vh t ) From this, we can get the predicted output value o of each sequence node t , combined with the loss function E t Get the actual output y of the sequence node t , determine whether it is a vehicle passing by, and track the identified vehicle; S32. Calculate the loss term of each node through the back propagation method. Suppose the loss term of the mth layer at time t is It is propagated along the layer and timeline to the previous layer and previous moment of the input layer and hidden layer, and the current layer of the long-term memory knowledge element, and we get and Assume n t is the weighted input of the system at the time series node t, that is: n t =Ux t +WC t +Xh t-1 h t =n t By The term is decomposed into Calculate the loss term α and use the back propagation rule to get the loss term at any time k The function of diag[x] is to create a diagonal matrix based on the x vector; S33. Calculate the gradient of the loss function E with respect to each weight matrix U, W, X, V; The gradient of the loss function E with respect to the weight matrix U is calculated by taking the loss term at time t With input x t Multiply, repeat the operation over the time series and sum the gradient of the obtained weight matrix U Similarly, the gradient of the loss function E with respect to the weight matrix W is calculated by taking the loss term at time t and the long-term memory knowledge element C at this moment t Multiply, repeat the operation over the time series and sum the gradient of the obtained weight matrix W Similarly, the gradient of the loss function E with respect to the weight matrix X is calculated by taking the loss term at time t and the hidden item h at the previous moment t-1 Multiply, repeat the operation over the time series and sum the gradient of the obtained weight matrix X The gradient of the loss function E with respect to the weight matrix V is: The gradient formulas are all summations, which may cause gradient vanishing or gradient exploding problems due to the length of the sequence; The RNN-LSTM algorithm in step S33 may have the problem of gradient vanishing or gradient exploding due to the length of data information. This method uses a GRU network structure to improve the RNN-LSTM algorithm, specifically including the following steps: S431. Based on the original RNN-LSTM, GRU retains the dual-input and dual-output structure and adds two internal gate structures, "update gate" and "reset gate". The hidden state h obtained by the previous sequence node transmission t-1 and the input item x of the current node t And the weight matrix Q, the activation function β to obtain the gate signals N, S, where N controls the update gate (renewgate) and S controls the reset gate (resetgate); the values ​​of N and S belong to the range of (0, 1); S432. By obtaining the "reset gate" gating signal S, the data is "reset": h′ t-1 =h t-1 *S The reset hidden state h′ t-1 The input item x of the current node t Combined with the function th, the value is controlled in the interval (-1, 1) to obtain h′ t : Using the obtained "update gate" gating signal N and h' t h for this node t Perform "update" and use the value of N to control the influence of the previous node on this node. The value of N belongs to the range of (0, 1). The closer the value is to 0, the smaller the influence is: h t =N(h t-1 -h′ t )+h′ t 。 2. According to claim 1, a smart energy-saving lamp pole group control method based on a recurrent neural network algorithm is characterized in that: The RNN-LSTM algorithm in step S31 can reduce computational complexity and improve computational efficiency through time parallel computation; the specific steps are as follows: S51. Hidden state h of the previous node t-1 and the input item x of the current node t Apply weight matrix: Ux t +Xh t-1 , directly solve in parallel to obtain the intermediate variable i of the forget gate t , input gate intermediate variable f t and z t , output gate intermediate variable o t ; S52. Forget gate intermediate variable i obtained by parallel calculation t and long-term memory knowledge element C t-1 Apply weight matrix: i t +WC t-1 , get the short-term memory forgetting element S t ; At the same time, the input gate intermediate variable f t and z t Added together, it is the intermediate variable c′ of the long-term memory knowledge element t ; S53. Forgetting elements of short-term memory t and the intermediate variable c′ of the long-term memory knowledge element t Apply weight matrix: c′ t +WS t , get long-term memory knowledge element C t And store it, then output the intermediate variable o t 、Long-term memory knowledge element C t Apply the weight matrix: o t +WC t , and use the activation function to activate the hidden state h t ; The subsequent activation function steps are the same as those in step S31.

3. According to claim 1, a smart energy-saving lamp pole group control method based on a recurrent neural network algorithm is characterized in that: The specific method for identifying and calculating the vehicle speed in step S3 is to use the headlights of the vehicle identified in the video as the base point, and calculate the average speed of the vehicle on the road section based on the time it takes for it to pass through two specified virtual sections, which is equivalent to the instantaneous speed of the vehicle at this point.

Images