A deep learning-based highway tunnel intelligent traffic control method and system

Abstract

The application relates to the technical field of traffic control, and discloses a highway tunnel intelligent traffic control method and system based on deep learning. The method comprises the following steps: acquiring a real-time video stream, obtaining an image sequence through frame extraction and denoising; analyzing the image to obtain vehicle contour, density and type distribution data; calculating regional traffic load according to the data to obtain a load index value. If the index exceeds a preset threshold, adjusting the lighting intensity of the corresponding region to generate an initial ventilation power suggestion; combining the suggestion and the index to dynamically allocate power distribution to obtain an initial power scheme. The scheme is optimized and put into operation to obtain equipment configuration; combining the configuration and the vehicle contour to monitor temporary air turbulence fluctuations caused by an electric motorcycle cluster. If the fluctuations exceed a preset threshold, fine-tuning resources to obtain a final power distribution scheme. The method can solve the problem of insufficient sensing accuracy in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of traffic control, and in particular to a highway tunnel intelligent traffic control method and system based on deep learning. BACKGROUND

[0002] At present, highway tunnel traffic management is an important part of the modern transportation system, and its core is to ensure safety, improve efficiency and optimize energy utilization. With the acceleration of urbanization and the surge in traffic flow, tunnels, as the throat of the transportation hub, bear a huge traffic pressure, and the intelligent management realized by deep learning directly affects road safety and traffic experience.

[0003] In one prior art, a combination scheme of coil detector and analog video monitoring is adopted, a set of induction coils is buried every 500 meters below the entrance, middle and exit lanes of the tunnel, the vehicle flow is counted through electromagnetic induction principle, the sampling frequency is 1 time per minute, an analog camera is installed every 300 meters at the top of the tunnel, manual monitoring of picture changes, lighting control according to fixed power parameters of tunnel section, based on historical data, the high-speed gear of the fan is forced to start for ventilation control during the daily peak period, and the low-speed operation is carried out at a fixed frequency of 3 times per hour during the remaining period.

[0004] However, the coil has a large counting error when the traffic is congested, and cannot distinguish the vehicle type, and has no recognition ability for special vehicles such as heavy trucks, when manually monitoring the picture changes, the recognition accuracy is greatly reduced at night or in bad weather due to the influence of alternating lights in the tunnel, vehicle shadows and water mist, and the vehicle density data cannot be output quantitatively, and in summary, the prior art has the problem of insufficient perception accuracy. SUMMARY

[0005] The present application provides a highway tunnel intelligent traffic control method and system based on deep learning to solve the problem of insufficient perception accuracy in the prior art.

[0006] In a first aspect, to solve the above technical problems, the present application provides a highway tunnel intelligent traffic control method based on deep learning, comprising:

[0007] Obtaining a real-time video stream, and performing frame extraction and denoising on the real-time video stream to obtain an image sequence;

[0008] Analyzing the vehicle contour and motion trajectory of the image sequence to obtain vehicle contour extraction results, vehicle density data and type distribution data;

[0009] Calculating the regional traffic load according to the vehicle density data and the type distribution data to obtain a load index value;

[0010] If the load index value is higher than a preset index threshold, the lighting device in the corresponding area is adjusted in power, lighting intensity data is determined, and an initial ventilation power suggestion value is generated according to the lighting intensity data;

[0011] The initial ventilation power suggestion value is combined with the load index value, power distribution resources are dynamically allocated, and an initial power distribution scheme is obtained;

[0012] The initial power distribution scheme is iteratively optimized to obtain an optimized power distribution scheme, and the optimized power distribution scheme is input into a device to run to obtain a device running configuration;

[0013] According to the device running configuration combined with the vehicle contour extraction result, temporary air turbulence fluctuations caused by the electric motorcycle cluster in the vehicle contour extraction result are dynamically monitored to obtain turbulence fluctuations;

[0014] If the turbulence fluctuations exceed a preset fluctuation threshold, the allocation resources are fine-tuned to obtain a final power distribution scheme.

[0015] In a second aspect, the present application provides a highway tunnel intelligent traffic control system based on deep learning, comprising:

[0016] A data acquisition module is configured to acquire real-time video streams, perform frame extraction and denoising on the real-time video streams, and obtain image sequences;

[0017] A data analysis module is configured to analyze vehicle contours and motion trajectories of the image sequences to obtain vehicle contour extraction results, vehicle density data, and type distribution data;

[0018] A data calculation module is configured to calculate regional traffic loads according to the vehicle density data and the type distribution data to obtain load index values;

[0019] A data adjustment module is configured to adjust the power of lighting devices in corresponding areas if the load index value is higher than a preset index threshold, determine lighting intensity data, and generate an initial ventilation power suggestion value according to the lighting intensity data;

[0020] A data distribution module is configured to combine the initial ventilation power suggestion value with the load index value, dynamically allocate power distribution resources, and obtain an initial power distribution scheme;

[0021] A data optimization module is configured to iteratively optimize the initial power distribution scheme to obtain an optimized power distribution scheme, input the optimized power distribution scheme into a device to run, and obtain a device running configuration;

[0022] A data detection module is configured to dynamically monitor temporary air turbulence fluctuations caused by the e-motorcycle cluster in the vehicle contour extraction result based on the device operation configuration in combination with the vehicle contour extraction result, and obtain turbulence fluctuations.

[0023] A data fine-tuning module is configured to fine-tune the allocated resources if the turbulence fluctuations exceed the preset fluctuation threshold, and obtain a final power allocation scheme.

[0024] In a third aspect, the present application further provides an electronic device, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the deep learning-based highway tunnel intelligent traffic control method according to any one of the above.

[0025] In a fourth aspect, the present application further provides a computer-readable storage medium comprising a stored computer program, wherein the computer program controls a device where the computer-readable storage medium is located to execute the deep learning-based highway tunnel intelligent traffic control method according to any one of the above when the computer program is running.

[0026] Compared with the prior art, the present application has the following beneficial effects:

[0027] (1) The present application synchronously collects and pre-processes multiple cameras, and uses a convolutional neural network to recognize contours and a light flow method to track trajectories, to provide a basis for precise recognition by aligning high-definition image sequences with timestamps, avoiding the misplacement problem of traditional multi-device data; the classification accuracy of vehicle types (sedans / trucks / e-motorcycle clusters) is improved (based on contour features and a pre-trained model), solving the limitation of traditional "inability to distinguish vehicle types"; pixel value variance verification ensures that only clear images are analyzed, avoiding recognition errors caused by noise in traditional analog videos, and the accuracy of vehicle density data quantification is greatly improved.

[0028] (2) The present application constructs a dynamic response chain through load index triggering lighting adjustment, dynamic allocation of power distribution, and reinforcement learning iterative optimization, based on real-time load index values and the quantized values of vehicle type weights instead of fixed time triggering control, to ensure that lighting / ventilation power matches actual traffic demand; the reinforcement learning model continuously optimizes the scheme based on historical data and real-time density deviation, avoiding the traditional "one-size-fits-all" mode, achieving dynamic balance of "load-resources", and effectively improving energy utilization.

[0029] (3) This invention achieves precise allocation by combining ventilation recommendations with dynamic load allocation and turbulence-triggered fine-tuning. Based on the "combined data value" (ventilation demand plus load index), resources are preferentially allocated to high-demand areas (such as dense traffic areas in the middle of tunnels), avoiding the inefficiency of traditional average allocation. For temporary turbulence caused by electric motorcycle clusters, the local ventilation power is finely adjusted in a directional manner to solve the problem of insufficient response of traditional solutions to "special cluster scenarios" and ensure the air circulation balance in narrow sections. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the intelligent traffic control method for highway tunnels based on deep learning provided in the first embodiment of the present invention;

[0031] Figure 2 This is a schematic diagram of the structure of a deep learning-based intelligent traffic control system for highway tunnels provided in the second embodiment of the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Reference Figure 1 The first embodiment of the present invention provides a deep learning-based intelligent traffic control method for highway tunnels, comprising the following steps:

[0034] S11, acquire the real-time video stream, and perform frame extraction and denoising on the real-time video stream to obtain an image sequence;

[0035] S12, Analyze the vehicle outline and motion trajectory of the image sequence to obtain vehicle outline extraction results, vehicle density data and type distribution data;

[0036] S13, calculate the regional traffic load based on the vehicle density data and the type distribution data to obtain the load index value;

[0037] S14, if the load index value is higher than the preset index threshold, the power of the lighting equipment in the corresponding area is increased and adjusted, the lighting intensity data is determined, and an initial ventilation power recommendation value is generated based on the lighting intensity data;

[0038] S15, combine the initial ventilation power recommendation value with the load index value, dynamically allocate power distribution resources, and obtain the initial power allocation scheme;

[0039] S16, iteratively optimize the initial power allocation scheme to obtain an optimized power allocation scheme, and input the optimized power allocation scheme into the device for operation to obtain the device operation configuration;

[0040] S17, Based on the device operation configuration and the vehicle contour extraction result, dynamically monitor the temporary air turbulence fluctuations caused by the electric motorcycle cluster in the vehicle contour extraction result to obtain the turbulence fluctuations.

[0041] S18, if the turbulence fluctuation exceeds the preset fluctuation threshold, the resource allocation is fine-tuned to obtain the final power allocation scheme.

[0042] In step S11, a real-time video stream is acquired, and frame extraction and denoising are performed on the real-time video stream to obtain an image sequence, including:

[0043] The timestamps of each video stream in the real-time video stream are aligned to obtain a video stream sequence with consistent time.

[0044] The video stream sequence is subjected to frame extraction and denoising to obtain an image sequence.

[0045] It should be noted that multiple cameras (e.g., 3 cameras) are deployed at key locations such as the tunnel entrance, middle section, and exit, each capturing a 1080p real-time video stream at 30 frames per second. The video streams are timestamped by adding a uniform millisecond-level timestamp (e.g., based on the tunnel control system's standard clock) to each frame to eliminate hardware clock deviations. After alignment, a "timestamp-video frame" sequence is generated, resulting in the corresponding first video stream sequence. To reduce data redundancy (the 30 frames per second video contains a large amount of repetitive information), keyframes are extracted at intervals of 1 frame every 5 frames. To address video noise caused by dim lighting and flickering lights within the tunnel, a 5x5 Gaussian kernel is used to smooth the image, with the standard deviation of the Gaussian kernel being... The value is set to 1.0, which was determined after frequency domain analysis of typical noise images inside the tunnel. This value can effectively suppress headlight flicker noise while preserving vehicle edge features. The 5x5 Gaussian kernel is a 5x5 matrix, and the weight of each pixel value is determined by the Gaussian function (…). The calculation generates the image and removes light noise.

[0046] Gaussian filtering works by weighted averaging of neighboring pixel values. For each frame in a temporally consistent video stream sequence, a 5x5 Gaussian kernel is used as a sliding window to cover each pixel in the image (e.g., starting from the top-left pixel, with a step size of 1 pixel). For the 25 pixels covered by the current window (including the center pixel and its four surrounding rings of pixels), the grayscale value of each pixel is multiplied by the weight corresponding to the Gaussian kernel position, and then summed to obtain a new pixel value. For example, if the center pixel has a grayscale value of 200 and a weight of 0.153, and the neighboring pixel above it has a grayscale value of 190 and a weight of 0.124, then the weighted sum at that position is (200 × 0.153) + (190 × 0.124) + ... (the weighted values ​​of the remaining 23 pixels). Random noise in the tunnel (such as a single bright spot pixel whose gray value suddenly increases to 255) will be "diluted" by the low weight of the surrounding normal pixels. However, the vehicle edges can still retain obvious edge features after weighting because the gray value of neighboring pixels changes continuously (such as the gray value difference between the vehicle body and the background is stable), ultimately generating a clear image sequence.

[0047] In step S12, the vehicle contours and motion trajectories of the image sequence are analyzed to obtain vehicle contour extraction results, vehicle density data, and type distribution data, including:

[0048] Calculate the pixel value variance of the image sequence. If the pixel value variance is greater than a preset sharpness threshold, then extract the vehicle contour for each frame of the image to obtain the vehicle contour extraction result.

[0049] Based on the vehicle contour extraction results, the vehicle type is identified, and type distribution data is obtained;

[0050] The motion trajectory of the image sequence is analyzed, and the speed and density distribution are calculated to obtain vehicle density data.

[0051] It should be noted that variance reflects the dispersion of pixel grayscale values. A large variance (e.g., significant difference in brightness between a vehicle and the background inside a tunnel) indicates rich image detail and high clarity; a small variance (e.g., overall blur caused by heavy fog) indicates poor image quality. For the image sequence obtained in step S11 (denoised keyframes), the variance of grayscale pixel values ​​is calculated frame by frame. The variance formula is F0. Where F represents variance, The grayscale value (0-255) of a single pixel. The average grayscale value of the entire frame is N, where N is the total number of pixels (e.g., 2,073,600 pixels for a 1080p image).

[0052] A preset sharpness threshold is established (this threshold can be selected by analyzing the variance range of historical images; analysis shows that when the variance > 100, the grayscale difference between the vehicle and the background is significant, allowing for stable contour extraction). If the variance of a frame exceeds this threshold, the frame is deemed sharp and usable, triggering vehicle contour extraction. For images that pass the sharpness check, Gaussian filtering is used to further smooth the image (reducing noise interference). Pixel gradients (horizontal / vertical directions) are calculated to locate edges. The pixel gradient is the rate of change of pixel grayscale in the horizontal (x-axis) and vertical (y-axis) directions, reflecting the intensity and direction of the edge. The Sobel operator (a discrete differential operator) is used to calculate the gradient. A horizontal Sobel kernel is convolved with the image to detect vertical edges (such as the left and right edges of the vehicle body) and calculate the horizontal gradient. ); using a vertical Sobel kernel convolved with the image, horizontal edges (such as the top and bottom edges of a car body) are detected, and the vertical gradient is calculated. ); gradient is A larger gradient value indicates a more distinct edge.

[0053] Effective edges are filtered using a dual thresholding method (high threshold suppresses weak edges, low threshold retains strong edges). The dual thresholds are the core parameters for filtering effective edges in Canny edge detection, used to distinguish between real and noisy edges. The high threshold suppresses weak edges, retaining only edges with significant intensity (such as strong contrast edges between a car body and the background); the low threshold retains pixels that are weaker but may belong to real edges, avoiding missing edge fragments. The high threshold is typically twice the low threshold. Let's say the high threshold is 100 and the low threshold is 50. Based on the gradient magnitude statistics of vehicle edges within the tunnel, the gradient magnitude of real vehicle edges is mostly greater than 80, while the gradient magnitude of noise (such as flickering lights) is mostly less than 50. Therefore, setting the high threshold to 100 can filter out most noise, and setting the low threshold to 50 can retain weak edge fragments connected to strong edges. The edge fragments are then connected to generate a complete vehicle outline (such as the rectangular outline of a car or the narrow, elongated outline of a motorcycle), and the vehicle outline extraction result is output (including outline coordinates, dimensions, etc.). If the threshold is not exceeded (e.g., due to water mist or excessively low light causing blurring), the frame is skipped or image enhancement (such as contrast stretching) is triggered.

[0054] Based on contour features (size, shape, and proportion), a pre-trained convolutional neural network (open-source YOLOv5 model) is used for classification. Fine-tuning is then performed using tunnel traffic image data. The feature vectors of the vehicle contours (such as length, width, and aspect ratio) are used as input. Feature differences are learned through training samples (labeled as "cars, trucks, electric motorcycle clusters," etc.). For example, the aspect ratio of car contours is approximately 0.3-0.4, trucks approximately 0.5-0.6, and electric motorcycle clusters approximately 0.7-0.8. The final output is type distribution data (e.g., "In a certain frame of the image, there are 15 cars, 3 trucks, and 8 electric motorcycle clusters").

[0055] For vehicle contours in consecutive frames, optical flow is used to calculate the motion trajectory. By analyzing the pixel displacement vectors of the same contour in adjacent frames, the direction and distance of vehicle movement are determined (e.g., at time t1, the contour center is at coordinates (x1, y1), and at time t2, it is at (x2, y2), with a displacement of...). Output the trajectory curve of each vehicle (including timestamp and location coordinates). Count the number of vehicles per unit time (e.g., 1 minute) and per unit tunnel length (e.g., 1 kilometer) to generate vehicle density data (e.g., "20 vehicles pass through the middle section of the tunnel in 1 minute, density is 20 vehicles / km").

[0056] In step S13, the regional traffic load is calculated based on the vehicle density data and the type distribution data to obtain a load index value, including:

[0057] By aligning the timestamps of the vehicle density data and the type distribution data, a time-aligned dataset is obtained;

[0058] Based on the time-aligned dataset, weights are assigned to each vehicle type to obtain vehicle type weights;

[0059] The regional traffic load is calculated based on the vehicle type weights to obtain the load index value.

[0060] It should be noted that the vehicle density data obtained through motion trajectory analysis in step S12 is timestamped based on the end time of trajectory statistics, and the type distribution data obtained through vehicle type identification in step S12 is timestamped based on the completion time of type classification. The same time synchronization benchmark (Network Time Protocol NTP) as in step S11 is used to calibrate the timestamps of the two types of data to the same time granularity (e.g., accurate to the second). For example, the density data at 10:01:00 and the type data at 10:01:02 are both aligned to the 1-minute window of 10:01:00-10:02:00 to ensure that the total number of vehicles in a certain period corresponds one-to-one with the type percentage in that period, generating a time-aligned dataset of "time window - total number of vehicles - number of types" (e.g., "10:01-10:02: 20 vehicles (15 cars, 5 trucks)").

[0061] Based on the differences in the impact of vehicles on tunnel load (emission intensity, volume, whether they cause turbulence, etc.), weights are assigned to each vehicle type: passenger cars = 1 (low emissions, small size, low ventilation / lighting requirements); trucks = 2 (high emissions, large size, ventilation requirements are twice that of passenger cars); electric motorcycle convoys = 1.5 (convoy driving easily causes air turbulence, requiring additional ventilation); buses = 3 (large passenger capacity, concentrated emissions, most significant load impact). The weights are based on historical data statistics (e.g., the rate of CO concentration increase when trucks pass is twice that of passenger cars) and tunnel design standards (specifications on the correlation between ventilation volume and vehicle type). A "vehicle type-weight" mapping table is generated (e.g., {passenger cars: 1, trucks: 2, electric motorcycle convoys: 1.5, buses: 3}), resulting in vehicle type weights; that is, determined comprehensively based on the actual emission data, volume proportion, and measured data of the impact on airflow disturbance of each type of vehicle in the tunnel.

[0062] The load index value comprehensively reflects the combined impact of vehicle quantity (density) and type (scale factor) on the regional load—the higher the value, the greater the demand for ventilation and lighting in the area (e.g., an index value of 25 requires more power distribution resources than an index value of 10). Based on the time-aligned dataset and vehicle type weights, the load index value is calculated by weighted summation. The load index value equals the number of vehicles of a certain type multiplied by the corresponding weight. For example, if the "10:01-10:02 time period" in the time-aligned dataset contains 15 cars (weight 1) and 5 trucks (weight 2), then the load index value = 15 × 1 + 5 × 2 = 15 + 10 = 25.

[0063] In step S14, if the load index value is higher than a preset index threshold, the power of the lighting equipment in the corresponding area is increased and adjusted. Lighting intensity data is determined, and an initial ventilation power recommendation value is generated based on the lighting intensity data, including:

[0064] Acquire ventilation equipment status data and air quality parameters;

[0065] If the load index value is higher than the preset index threshold, the power of the lighting equipment in the corresponding area will be increased and adjusted to determine the lighting intensity data;

[0066] Based on the lighting intensity data, the air quality parameters are classified to obtain a classification dataset;

[0067] Based on the classification dataset, the power of the ventilation equipment is optimized using a preset lighting and ventilation linkage coefficient to obtain an initial recommended ventilation power value.

[0068] It should be noted that the ventilation equipment status data collected through the tunnel equipment monitoring system includes the current operating power of the fan (e.g., 50 kW), speed (high / low speed), cumulative running time, and fault codes (0 for no faults), reflecting the current controllability of the equipment (e.g., whether the fan is at full load). Air quality parameters, including CO concentration (e.g., 100 ppm), visibility (e.g., 90%), and PM2.5 concentration (e.g., 50 μg / m³), are collected through sensors deployed in the tunnel (CO sensor, visibility meter, particulate matter sensor), reflecting the current environmental quality. All data are appended with a timestamp consistent with the "load index value" in step S13 (e.g., "10:01-10:02 time period") to ensure that the "load-equipment-environment" data matches in the time dimension.

[0069] The lighting and ventilation linkage coefficient is generated by fitting historical data. For every 100 lux increase in tunnel lighting intensity, the ventilation system should simultaneously increase its air exchange rate by 15%. Verification using three years of operational data from a 2km long, four-lane tunnel showed that when lighting intensity increased from 300 lux to 500 lux, CO concentration rose by an average of 12-18%, requiring a 15% increase in ventilation power to maintain CO concentration ≤100ppm. When visibility is <80% or rainfall is >5mm / h, the linkage coefficient automatically adjusts to 20%.

[0070] The preset threshold values ​​are set based on the tunnel's design capacity and historical load data. According to the simulation results of the tunnel traffic flow simulation software VISSIM, when the average daily traffic flow exceeds 70% of the design capacity, the load index value is approximately 25, at which point the load control mechanism needs to be activated. If the load index value (e.g., 25) is higher than this threshold, it is determined to be a "high load area," and the lighting power needs to be increased. Adjustments are made according to the extent to which the load index value exceeds the threshold. For example, a load index value of 20-30 corresponds to "medium-high load," and the lighting intensity is increased from the base value (e.g., 300 lux) to 500 lux; above 30 corresponds to "high load," and the intensity is increased to 700 lux (according to tunnel lighting design specifications, to ensure clear vehicle visibility). Adjustments are only made to areas exceeding the load limit (e.g., the middle section of the tunnel), while the original power is maintained in non-exceeding areas to avoid energy waste, and the lighting intensity data is output (e.g., "500 lux in the middle section, 300 lux at the entrance / exit").

[0071] Lighting intensity data is correlated with air quality parameters by region to form a correlated dataset of "region-lighting intensity-air quality" (e.g., "Mid-section: 500 lux, CO 800 ppm, visibility 90%)). Air quality parameters are classified using a classification model employing a random forest algorithm. The training data includes 100,000 historical air quality samples (CO concentration, visibility), achieving an accuracy of 92% after 5-fold cross-validation. Pre-defined classification criteria are used, such as "Excellent" for CO concentration < 500 ppm and visibility > 95%; "Medium" for 500-800 ppm and visibility 85%-95%; and "Poor" for > 800 ppm or visibility < 85%. The output dataset is a classified dataset, such as "Mid-section: Lighting 500 lux, Air Quality Category 'Medium'".

[0072] Based on the classification dataset and combined with ventilation equipment status data (e.g., current power 50 kW, maximum increase to 80 kW), a mapping relationship of "lighting intensity - air quality category - ventilation power" is established. For example, if the classification is "excellent," it is recommended that lighting be 300 lux → ventilation power 40 kW; if the classification is "medium," it is recommended that lighting be 500 lux → ventilation power 60 kW (higher than the current 50 kW to improve air circulation); if the classification is "poor," it is recommended that lighting be 700 lux → ventilation power 75 kW (prioritizing air quality). Based on the above mapping, ventilation power recommendations adapted to the current load and environment are generated (e.g., "adjust the power of the mid-section fan to 60 kW").

[0073] In step S15, the initial ventilation power recommendation value is combined with the load index value to dynamically allocate power distribution resources, resulting in an initial power allocation scheme, including:

[0074] The initial recommended ventilation power value is combined with the load index value to obtain the combined data value;

[0075] If the combined data value is higher than the preset combined value threshold, a power distribution resource allocation instruction set is generated;

[0076] Based on the power distribution resource allocation instruction set, the ventilation equipment is allocated resources to obtain an initial power allocation scheme.

[0077] It should be noted that the initial ventilation power recommendation is based on air quality classification and lighting intensity (e.g., "60 kW in the middle of the tunnel"); the load index value is a quantitative value reflecting the regional traffic load (e.g., 25). A weighted summation algorithm is used to combine the two. The combined data value is equal to the product of ventilation weight multiplied by the initial ventilation power recommendation value, plus the product of load weight multiplied by the load index value. The ventilation weight and load weight are set according to the tunnel resource priority (e.g., if ventilation equipment has a greater impact on safety, the ventilation weight is taken as 0.6, and the load weight as 0.4). For example, the initial ventilation power recommendation value is 60 kW, divided by the base unit is 60, the load index value is 25, and the combined data value is 0.6 × 60 + 0.4 × 25 = 36 + 10 = 46.

[0078] Based on the tunnel power distribution system capacity and historical load data, a combined threshold value is set at 40. Exceeding this value indicates that the power distribution system needs dynamic adjustment to avoid overload. This threshold value is used to determine whether power distribution resources need to be adjusted. This threshold value reflects the "critical value of power distribution pressure after the current ventilation demand and traffic load are superimposed". Exceeding the threshold value indicates that the existing power distribution scheme is insufficient and dynamic adjustment is required. If the combined data value (e.g., 46) is greater than the preset combined threshold value (40), a "power distribution resource allocation instruction set" is generated, which includes the target area (e.g., "tunnel middle section"); equipment type (e.g., "ventilation fan unit 1#, 2#"); allocated power (e.g., "each fan is increased to 30 kilowatts, total power 60 kilowatts"); execution time (e.g., "effective immediately, lasting until the combined value is lower than the threshold value").

[0079] Based on the area and equipment information in the instruction set, power distribution resources are allocated according to priority, prioritizing ventilation equipment in high-load areas (such as the middle section of a tunnel) and then considering other areas, ensuring that resources are tilted towards the "core areas where combined data values ​​exceed the standard". The plan is presented in the form of a table or instruction list, clearly specifying the power distribution details of each device, and outputting the initial power allocation plan.

[0080] In step S16, the initial power allocation scheme is iteratively optimized to obtain an optimized power allocation scheme. This optimized power allocation scheme is then input into the device for operation, resulting in the device's operating configuration, including:

[0081] The distribution deviation is obtained by comparing the vehicle density data with a preset density standard;

[0082] When the distribution deviation exceeds a preset deviation threshold, the initial power allocation scheme is input into the trained reinforcement learning model for iterative optimization to obtain an optimized power allocation scheme.

[0083] The optimized power allocation scheme is input into the device for operation, resulting in the device's operating configuration.

[0084] It should be noted that the preset density standard can be based on the tunnel's design capacity and historical peak data to set a "reasonable density range" (e.g., "normal density of 10-15 vehicles / km in the middle section, 8-12 vehicles / km at the entrance"), reflecting the differences in traffic carrying capacity in different areas. The deviation between the real-time density and the preset standard is calculated by dividing the difference between the real-time vehicle density and the preset density standard value by the quotient of the preset density standard value and multiplying by 100%. For example, if the real-time density in the middle section of the tunnel is 20 vehicles / km and the preset standard value is 15 vehicles / km, the distribution deviation is approximately 33.3% (|(20-15) / 15|×100%), indicating that the density in this area is significantly higher than the reasonable range. The preset deviation threshold is set based on the tunnel's operational stability requirements. According to experimental data, when the distribution deviation exceeds 10%, the adaptability of the initial power allocation scheme drops below 85%, which is used to determine whether optimization is needed. If the distribution deviation (e.g., 33.3%) is greater than the threshold (10%), it indicates that the initial power allocation scheme is not adapted to the current density distribution (e.g., the density in the middle section is too high but the ventilation power is insufficient), and optimization needs to be initiated.

[0085] The reinforcement learning model adopts a deep Q-network (DQN) architecture. The state space includes vehicle density data (vehicles / km), equipment operating power (kW), and air quality parameters (CO concentration, visibility). The action space is the power adjustment amount (-10% to +20%). The reward function is designed as a weighted sum of energy consumption reduction rate (weight 0.3) and air quality improvement rate (weight 0.7). The model was trained using tunnel operation data from the past 6 months (approximately 100,000 samples) over a 500-epoch period with a learning rate of 0.001, a batch size of 32, and an experience replay pool capacity of 10,000. The goal was to minimize energy consumption and maximize environmental quality (e.g., CO concentration ≤ 800 ppm). An initial power allocation scheme (e.g., "mid-section ventilation 60 kW") was input into the model, and the model iteratively optimized its parameters based on real-time density deviation (33.3%). For example, the mid-section ventilation power was increased to 70 kW to match the high density, while the inlet redundant power was reduced (from 20 kW to 15 kW). Through multiple iterations (e.g., 5-10 times), an "optimized power allocation scheme" (e.g., "mid-section 70 kW, inlet 15 kW") was generated.

[0086] The optimized power allocation scheme is transformed into the operating parameters of specific equipment, forming an "equipment operation configuration", which includes: equipment identification (such as "ventilation fan 1# (middle section), 2# (inlet)"); operating power (such as "1#: 70kW, 2#: 15kW"); start / adjustment time (such as "execute immediately, continue until distribution deviation <10%)". The configuration is then distributed to the equipment through the tunnel industrial control system, and the equipment operating status (such as the actual power and current of the fan) is collected in real time to ensure that the configuration is implemented.

[0087] In step S17, based on the device operating configuration and the vehicle contour extraction results, the temporary air turbulence fluctuations caused by the electric motorcycle cluster in the vehicle contour extraction results are dynamically monitored to obtain the turbulence fluctuations, including:

[0088] The device operation configuration and vehicle contour extraction results are time-stamped and aligned with tunnel sections to obtain a contour feature set;

[0089] Real-time contour data of the electric motorcycle cluster is extracted from the contour feature set, and the temporary air turbulence fluctuations caused by the electric motorcycle cluster are dynamically monitored to obtain the turbulence fluctuations.

[0090] It should be noted that the equipment operation configuration includes the real-time operating parameters of the ventilation equipment in each tunnel section (such as "mid-section fan power 70kW, air supply direction forward"), operating status (such as "normal operation"), etc., with timestamps (accurate to the second) and section identifiers (such as "tunnel mid-section K2+100 to K2+300"); the vehicle contour extraction results include the contour coordinates, quantity, and spatial distribution of the electric motorcycle cluster (such as "electric motorcycle cluster at mid-section K2+150, a total of 20 vehicles"), etc., also with timestamps and section identifiers. Using the same Network Time Protocol (NTP) as in step S11, the timestamps of both types of data are calibrated to the same millisecond precision (e.g., "t=10:02:05.123") to ensure that the "equipment operating status" and "electric motorcycle cluster dynamics" are associated at the same time. Tunnel sections are aligned by matching the "segment identifier" of the equipment operating configuration with the "spatial location" of the vehicle contour extraction results according to the preset physical segments of the tunnel (e.g., every 200 meters is a segment) (e.g., both correspond to "middle section K2+100 to K2+300"), ensuring that they are consistent in spatial range. After alignment, a multi-dimensional feature set of "time-segment-equipment status-electric motorcycle cluster contour" is generated, for example, "t=10:02:05.123, middle section K2+100 to K2+300, fan power 70kW, electric motorcycle cluster of 20 vehicles (dense contours, average spacing 1.5 meters)".

[0091] Real-time contour data of electric motorcycle clusters are extracted, and key cluster features are selected and quantified from the contour feature set, including: cluster density, the number of electric motorcycle clusters in a unit segment (e.g., "20 vehicles / 200-meter segment"); spatial distribution, degree of aggregation (e.g., "dense, 90% of vehicles are concentrated within 100 meters"); motion state, average speed (e.g., "30 km / h"), and driving direction (e.g., "driving in the same direction"). These features directly reflect the cluster's potential to disturb airflow (because electric motorcycle clusters form a dense group of moving obstacles when driving, which significantly disturbs airflow more than dispersed vehicles). Combining equipment operation configuration (such as the basic airflow field formed by ventilation fans) and real-time sensor data (local airflow changes collected by wind speed sensors and air pressure sensors), the system monitors and calculates turbulence intensity (such as "0.3Pa fluctuation amplitude") and influence range (such as "spreading 50 meters along the driving direction") using a hybrid model that integrates CFD and machine learning. Sensor verification compares the model calculation results with the "wind speed pulsation frequency" and "air pressure change amplitude" measured by the wind speed sensor, and corrects the model parameters (such as adjusting the "equivalent flow resistance coefficient" of the cluster on the airflow). Finally, the system obtains the quantified "turbulence fluctuation" result (such as "fluctuation amplitude 0.35Pa, duration 15 seconds, affecting the middle section K2+120 to K2+180").

[0092] A hybrid model integrating computational fluid dynamics (CFD) and machine learning is used to simulate tunnel airflow based on the Navier-Stokes equations. Inputting electric motorcycle cluster characteristics (density, speed) and fan operating parameters, the theoretical turbulence intensity distribution is calculated. An LSTM network is then used to correct the CFD results. Inputting historical turbulence monitoring data (air pressure fluctuations collected by wind speed sensors) and real-time cluster characteristics, a pre-trained CFD module is used. By adjusting the perturbation source terms, the mean square error (MSE) between theoretical and measured values ​​is ≤0.05 Pa². Then, with the CFD parameters fixed, its output is used as LSTM input. The LSTM is trained with the label "measured turbulence value - CFD theoretical value," with the optimization objective being the minimization of MSE. Joint fine-tuning is performed, synchronously updating the CFD perturbation coefficients and LSTM weights, until the test set MSE ≤0.02 Pa² (accuracy meets tunnel safety monitoring requirements). The corrected turbulence fluctuation values ​​are output (e.g., "fluctuation amplitude 0.35 Pa, duration 15 seconds").

[0093] It should be noted that the CFD turbulence model selected is the standard k-ε model, based on the Navier-Stokes equations, and a 1:1 numerical model of the tunnel is established using open-source CFD software (such as Open FOAM). The mesh is a polyhedral mesh, with local refinement at the fan outlet and vehicle driving area, with a mesh size of 0.5m × 0.5m and a mesh size of 1m × 1m in other areas. The inlet is a velocity inlet (converted according to the fan power), and the outlet is a pressure outlet. The density, speed, and spatial distribution of different electric motorcycle clusters are used as moving obstacle source terms for input. By adjusting the turbulence model parameters (such as the constants in the k-ε model), the simulation results are made consistent with the historical measured data of a small number of high-precision sensors (such as ultrasonic anemometers), with a mean square error (MSE) ≤ 0.05 Pa². High-fidelity flow field data is output, and the turbulence intensity (such as turbulent kinetic energy k) of key monitoring points is extracted.

[0094] It is worth noting that turbulence intensity and real-time cluster characteristics (tunnel operation data from the past 6 months, including the density, speed, spatial distribution of electric motorcycle clusters, and corresponding sensor data such as wind speed and air pressure, totaling approximately 100,000 samples) were input into the LSTM. The LSTM network structure was a 2-layer LSTM (64 units per layer). Due to the short-term dependence of the turbulence time series, the 2-layer structure can capture more complex nonlinear features. The 64-unit structure was chosen after balancing model capacity and the risk of overfitting, because the amount of error data used to train the LSTM is much smaller than the original image data. Ad was selected. The AM optimizer, Adam (lr=0.0005), has a learning rate slightly lower than that of typical tasks (e.g., 0.001) because the learning target is a correction term with relatively small values ​​and variations, requiring more refined and stable updates. The sliding window size is 10. The loss function is the mean squared error (MSE). The LSTM network is fine-tuned with the goal of minimizing the final prediction error, using k-fold cross-validation (k=5). The validation results show that the model's prediction accuracy reaches over 90%, and the mean squared error (MSE) is ≤0.02 Pa², which meets the requirements for tunnel safety monitoring.

[0095] In step S18, if the turbulent fluctuations exceed a preset fluctuation threshold, the resource allocation is fine-tuned to obtain the final power allocation scheme.

[0096] It should be noted that turbulent fluctuations are a quantitative indicator of air disturbances caused by the electric motorcycle cluster (such as "fluctuation amplitude 0.6 Pa, affecting the middle section of the tunnel from K2+120 to K2+180"), which includes information such as fluctuation intensity, duration, and affected section;

[0097] The preset fluctuation threshold is based on the tunnel safety standard (0.5 Pa). This value has been verified through historical data. When the turbulence fluctuation is ≤0.5 Pa, it has no significant impact on driving visibility and air quality. If it exceeds this threshold, it may lead to decreased visibility or local airflow turbulence, requiring intervention. If the turbulence fluctuation (e.g., 0.6 Pa) is greater than the preset fluctuation threshold (0.5 Pa), it is judged as "turbulence exceeding the standard," triggering the power distribution resource fine-tuning mechanism. For specific sections affected by turbulence (e.g., the middle section from K2+120 to K2+180), combined with the real-time contour data of the electric motorcycle cluster (e.g., "20 vehicles clustered, average speed 30 km / h") and the current equipment operating configuration (e.g., "the current power of the fan in this section is 70 kW"), the fine-tuning direction is determined according to the core logic. That is, the turbulence is caused by the movement of the cluster, and the disturbance needs to be offset by increasing the local ventilation power. The higher the cluster density and the faster the speed, the greater the fine-tuning amplitude (e.g., if the cluster density increases by 50%, the ventilation power will increase by 15% accordingly).

[0098] The "targeted incremental adjustment" method is adopted, which only fine-tunes the power of ventilation equipment in the turbulence-affected section to avoid energy waste caused by the adjustment of the entire area. For example, if the current section's fan power is 70kW and the turbulence fluctuation is 0.6Pa (exceeding the threshold by 0.1Pa), according to the rule of "increasing the power by 5% for every 0.1Pa exceeding the threshold" (this rule is based on the statistical analysis of historical operating data, and when the turbulence fluctuation exceeds the threshold by 0.1Pa, increasing the ventilation power by 5% can suppress the turbulence to a safe range within 30 seconds in 90% of cases), the fine-tuned power = 70 × (1 + 5%) = 73.5kW. The fine-tuning parameters are recorded simultaneously, including the target equipment (such as "mid-section fans 1# and 2#"), the adjusted power (73.5kW), and the execution time (such as "continue until the turbulence fluctuation is ≤0.5Pa"). The fine-tuned parameters are integrated into a "final power allocation scheme", which clearly defines the section division, such as "the middle section of the tunnel from K2+120 to K2+180"; equipment configuration, such as "fan 1#: 73.5kW, fan 2#: 73.5kW"; and related conditions, such as "when the number of electric motorcycles in this section is less than 10 or the turbulence fluctuation is less than or equal to 0.5Pa, the power is restored to the optimized power of 70kW".

[0099] Preferably, after fine-tuning, the turbulence fluctuation value is continuously monitored. If it does not drop below the threshold within a set time (e.g., 30 seconds), the fine-tuning logic is restarted to gradually increase the ventilation power until the turbulence stabilizes. The effect of each fine-tuning is recorded for subsequent model optimization.

[0100] In summary, this invention achieves high-definition image sequences with timestamp alignment through simultaneous acquisition from multiple cameras, noise reduction preprocessing, convolutional neural network contour recognition, and optical flow trajectory tracking, providing a foundation for accurate recognition and avoiding the data misalignment problem of traditional multi-device systems. The classification accuracy of vehicle types (cars / trucks / electric motorcycles) is improved (based on contour features and a pre-trained model), overcoming the limitation of traditional methods that "cannot distinguish vehicle types." Pixel value variance verification ensures that only clear images are used for analysis, avoiding recognition errors caused by noise in traditional analog videos, and significantly improving the accuracy of vehicle density data quantification. This invention also constructs a dynamic response chain through load index-triggered lighting adjustments, dynamic power distribution allocation, and reinforcement learning iterative optimization. Based on real-time load index values, it integrates quantified values ​​of vehicle type weights rather than fixed-time triggering adjustments, ensuring that lighting / ventilation power matches actual traffic demand. The reinforcement learning model continuously optimizes the scheme by combining historical data and real-time density deviations, avoiding the traditional "one-size-fits-all" approach and achieving a dynamic balance between "load and resources," effectively improving energy utilization. This invention achieves precise allocation by combining ventilation recommendations with dynamic load distribution and turbulence-triggered fine-tuning. Based on "combined data values" (ventilation demand plus load index), resources are prioritized for allocation to high-demand areas (such as densely trafficked areas in the middle of tunnels), avoiding the inefficiency of traditional average allocation. For temporary turbulence caused by electric motorcycle clusters, the local ventilation power is fine-tuned in a targeted manner, solving the problem of insufficient response of traditional solutions to "special cluster scenarios" and ensuring air circulation balance in narrow passages.

[0101] Reference Figure 2 The second embodiment of the present invention provides a deep learning-based intelligent traffic control system for highway tunnels, comprising:

[0102] The data acquisition module is used to acquire real-time video streams and perform frame extraction and noise reduction on the real-time video streams to obtain image sequences;

[0103] The data analysis module is used to analyze the vehicle contours and motion trajectories of the image sequence to obtain vehicle contour extraction results, vehicle density data, and type distribution data.

[0104] The data calculation module is used to calculate the regional traffic load based on the vehicle density data and the type distribution data, and obtain the load index value.

[0105] The data adjustment module is used to adjust the power of the lighting equipment in the corresponding area if the load index value is higher than the preset index threshold, determine the lighting intensity data, and generate an initial ventilation power recommendation value based on the lighting intensity data.

[0106] The data allocation module is used to combine the initial ventilation power recommendation value with the load index value to dynamically allocate power distribution resources and obtain an initial power allocation scheme.

[0107] The data optimization module is used to iteratively optimize the initial power allocation scheme to obtain an optimized power allocation scheme, and input the optimized power allocation scheme into the device for operation to obtain the device operation configuration;

[0108] The data detection module is used to dynamically monitor the temporary air turbulence fluctuations caused by the electric motorcycle cluster in the vehicle contour extraction results based on the device's operating configuration and the vehicle contour extraction results, and obtain the turbulence fluctuations.

[0109] The data fine-tuning module is used to fine-tune the resource allocation if the turbulence fluctuation exceeds a preset fluctuation threshold, so as to obtain the final power allocation scheme.

[0110] It should be noted that the deep learning-based intelligent traffic control system for highway tunnels provided in this embodiment of the invention is used to execute all the process steps of the deep learning-based intelligent traffic control method for highway tunnels described in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0111] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a deep learning-based intelligent traffic control program for highway tunnels. When the processor executes the computer program, it implements the steps described in the various deep learning-based intelligent traffic control method embodiments for highway tunnels, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above-described device embodiments, such as the data fine-tuning module.

[0112] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0113] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0114] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting all parts of the electronic device via various interfaces and lines.

[0115] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0116] If the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0117] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0118] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A deep learning-based intelligent traffic control method for highway tunnels, characterized in that, include: Acquire a real-time video stream, and perform frame extraction and denoising on the real-time video stream to obtain an image sequence; The vehicle contours and motion trajectories of the image sequence are analyzed to obtain vehicle contour extraction results, vehicle density data, and type distribution data; The regional traffic load is calculated based on the vehicle density data and the type distribution data to obtain the load index value; If the load index value is higher than the preset index threshold, the power of the lighting equipment in the corresponding area is increased and adjusted, the lighting intensity data is determined, and an initial ventilation power recommendation value is generated based on the lighting intensity data. By combining the initial recommended ventilation power value with the load index value, power distribution resources are dynamically allocated to obtain an initial power allocation scheme. The initial power allocation scheme is iteratively optimized to obtain an optimized power allocation scheme. The optimized power allocation scheme is then input into the device for operation to obtain the device operation configuration. Based on the device's operating configuration and the vehicle contour extraction results, the system dynamically monitors the temporary air turbulence fluctuations caused by the electric motorcycle clusters in the vehicle contour extraction results to obtain the turbulence fluctuations. If the turbulence fluctuations exceed a preset fluctuation threshold, the resource allocation is fine-tuned to obtain the final power allocation scheme; The process of generating an initial ventilation power recommendation value based on the lighting intensity data includes: Acquire ventilation equipment status data and air quality parameters; Lighting intensity data and air quality parameters are correlated by region to form a correlated dataset of "region-lighting intensity-air quality"; Based on the lighting intensity data, the air quality parameters are classified to obtain a classification dataset; Based on the classification dataset and ventilation equipment status data, a mapping relationship of "lighting intensity - air quality category - ventilation power" is established to generate an initial ventilation power recommendation value that is adapted to the current load and environment. The initial ventilation power recommendation value is a ventilation equipment power recommendation generated based on air quality classification and lighting intensity. The initial power allocation scheme is obtained by combining the recommended initial ventilation power value with the load index value to dynamically allocate power distribution resources, including: The initial recommended ventilation power value is combined with the load index value to obtain the combined data value; If the combined data value is higher than the preset combined value threshold, a power distribution resource allocation instruction set is generated; Based on the power distribution resource allocation instruction set, the ventilation equipment is allocated resources to obtain an initial power allocation scheme; The initial power allocation scheme is iteratively optimized to obtain an optimized power allocation scheme, including: The distribution deviation is obtained by comparing the vehicle density data with a preset density standard; When the distribution deviation exceeds a preset deviation threshold, the initial power allocation scheme is input into the trained reinforcement learning model for iterative optimization to obtain an optimized power allocation scheme.

2. The intelligent traffic control method for highway tunnels based on deep learning according to claim 1, characterized in that, The process of acquiring a real-time video stream and performing frame extraction and denoising on the real-time video stream to obtain an image sequence includes: The timestamps of each video stream in the real-time video stream are aligned to obtain a video stream sequence with consistent time. The video stream sequence is subjected to frame extraction and denoising to obtain an image sequence.

3. The intelligent traffic control method for highway tunnels based on deep learning according to claim 1, characterized in that, The analysis of the vehicle contours and motion trajectories in the image sequence yields vehicle contour extraction results, vehicle density data, and type distribution data, including: Calculate the pixel value variance of the image sequence. If the pixel value variance is greater than a preset sharpness threshold, then perform vehicle contour extraction on each frame of the image to obtain the vehicle contour extraction result. Based on the vehicle contour extraction results, the vehicle type is identified, and type distribution data is obtained; The motion trajectory of the image sequence is analyzed, and the speed and density distribution are calculated to obtain vehicle density data.

4. The intelligent traffic control method for highway tunnels based on deep learning according to claim 1, characterized in that, The step of calculating the regional traffic load based on the vehicle density data and the type distribution data to obtain the load index value includes: By aligning the timestamps of the vehicle density data and the type distribution data, a time-aligned dataset is obtained; Based on the time-aligned dataset, weights are assigned to each vehicle type to obtain vehicle type weights; The regional traffic load is calculated based on the vehicle type weights to obtain the load index value.

5. The intelligent traffic control method for highway tunnels based on deep learning according to claim 1, characterized in that, The step involves dynamically monitoring temporary air turbulence fluctuations caused by the electric motorcycle cluster in the vehicle contour extraction results, based on the device's operating configuration and the vehicle contour extraction results, to obtain the turbulence fluctuations, including: The device operation configuration and vehicle contour extraction results are time-stamped and aligned with tunnel sections to obtain a contour feature set; Real-time contour data of the electric motorcycle cluster is extracted from the contour feature set, and the temporary air turbulence fluctuations caused by the electric motorcycle cluster are dynamically monitored to obtain the turbulence fluctuations.

6. A deep learning-based intelligent traffic control system for highway tunnels, characterized in that, The method for implementing the deep learning-based intelligent traffic control method for highway tunnels as described in any one of claims 1 to 5 includes: The data acquisition module is used to acquire real-time video streams and perform frame extraction and noise reduction on the real-time video streams to obtain image sequences; The data analysis module is used to analyze the vehicle contours and motion trajectories of the image sequence to obtain vehicle contour extraction results, vehicle density data, and type distribution data. The data calculation module is used to calculate the regional traffic load based on the vehicle density data and the type distribution data, and obtain the load index value. The data adjustment module is used to adjust the power of the lighting equipment in the corresponding area if the load index value is higher than the preset index threshold, determine the lighting intensity data, and generate an initial ventilation power recommendation value based on the lighting intensity data. The data allocation module is used to combine the initial ventilation power recommendation value with the load index value to dynamically allocate power distribution resources and obtain an initial power allocation scheme. The data optimization module is used to iteratively optimize the initial power allocation scheme to obtain an optimized power allocation scheme, and input the optimized power allocation scheme into the device for operation to obtain the device operation configuration; The data detection module is used to dynamically monitor the temporary air turbulence fluctuations caused by the electric motorcycle cluster in the vehicle contour extraction results based on the device's operating configuration and the vehicle contour extraction results, and obtain the turbulence fluctuations. The data fine-tuning module is used to fine-tune the resource allocation if the turbulence fluctuation exceeds a preset fluctuation threshold, so as to obtain the final power allocation scheme.

Images