Tunnel portal dynamic light and signal control system based on vision algorithm

Abstract

The invention discloses a tunnel entrance dynamic light and signal control system based on a visual algorithm, relates to the technical field of intelligent manufacturing control, and is used for solving the problem of poor light control when the visual field of a tunnel entrance suddenly changes. According to the method, a dynamic sensing mechanism based on visual feature understanding and multi-strategy linkage is constructed, continuous monitoring and intelligent control over the driving visual adaptation state of the tunnel portal are achieved, multi-dimensional visual features are extracted through image collection and preprocessing, visual risk scores are generated, visual field sudden change and traffic abnormity are recognized in combination with a sliding window, and the driving visual adaptation state of the tunnel portal is obtained. A risk response instruction is output, illumination adjustment and signal linkage control are driven, then vehicle behavior continuity, parking probability and blind area shielding factors are fused, speed limiting, guiding or warning signals are dynamically issued, and visual model fine adjustment and strategy updating are performed based on execution feedback. And finally, illumination and signal adaptive optimization under cross-time and multi-environment conditions is realized in combination with a control scene library, and the traffic safety and control intelligence level are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing control technology, and more specifically, to a tunnel entrance dynamic lighting and signal control system based on a vision algorithm. Background Art

[0002] At present, the tunnel traffic system is an important part of the urban trunk and highway transportation infrastructure. Its operational safety and traffic efficiency directly affect the operation quality of the entire road. In most tunnel entrance areas, due to the obvious brightness difference between external natural light and internal tunnel lighting, especially on sunny days, backlight, early morning and evening time, drivers are very likely to experience visual adaptation disorders when entering the tunnel, resulting in instantaneous blurred vision or even a "black hole effect", which has become a high incidence of traffic accidents.

[0003] Deficiencies in existing technologies: When a vehicle enters the tunnel entrance section at high speed from bright outdoor conditions, the traditional system makes judgments based solely on the global illumination value obtained by the brightness sensor, failing to comprehensively analyze visual influencing factors such as vehicle density, traffic speed, incident light angle, and reflective interference. This results in the inability to adjust the light intensity at the entrance section in a timely manner, or the adjustment change curve does not match the actual driving visual state. Especially in complex light environments such as backlighting, strong glass reflection, and the junction of yin and yang, it is difficult to issue auxiliary signals in a timely manner or increase the lighting level in the entrance area, directly triggering a chain reaction from the driver's visual maladaptation to slow operation and then increased risk, thereby affecting the safe driving status of vehicles at the tunnel entrance. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, there is a solution as follows to solve the problem of poor lighting control due to sudden changes in the field of view at the tunnel entrance in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions: The tunnel entrance dynamic lighting and signal control system based on visual algorithms includes a tunnel entrance image acquisition module, a response instruction generation module, an illumination adjustment module, a control command generation module, an update input module, and a joint control module. Each module is connected by signals. The tunnel entrance image acquisition module collects a continuous sequence of image frames from the tunnel entrance area, performs defogging, contrast enhancement, and lane normalization processing, and extracts visual features from the image frames. The response instruction generation module performs tensor encoding on the extracted visual features and calculates the driving visual adaptation risk score using the built-in visual adaptation model. It then uses a sliding time window to detect short-term visual field mutation areas and speed reduction sections, and outputs a set of risk response instruction candidates. The illumination adjustment module builds a lighting control state map based on the risk response instruction candidate set, extracts the matching relationship between traffic flow characteristics and light gradients, and maps and generates a set of entrance section lighting illumination adjustment parameters; The control command generation module simultaneously constructs signal linkage trigger factors, combining the continuity of the tunnel entry behavior, the probability of the vehicle stopping, and the degree of visual blind spot obstruction to determine whether to generate speed limit, guidance, or warning signal control commands; The update input module sends the lighting illumination adjustment parameter group and signal control command to the on-site control device for execution. At the same time, it collects the execution results and image change sequence, extracts the lighting response deviation and behavior trajectory offset as the update input of the visual adaptation model; The joint control module, based on behavioral feedback results, adopts a fine-tuning mechanism to update the brightness response mapping relationship and linkage strategy rules in the visual adaptation model, and adaptively optimizes the joint control of lights and signals in different time periods and environments.

[0006] Furthermore, a continuous sequence of image frames of the tunnel entrance area is collected and defogging, contrast enhancement, and lane normalization are performed to extract visual features of the image frames, including: Dehazing the image frame is performed using dark channel prior and histogram equalization algorithm; Performing contrast enhancement on image frames using contrast-limited adaptive histogram equalization method; Based on image perspective correction and lane template matching algorithms, lane normalization is performed on image frames to convert lane structures from different perspectives into a unified spatial reference system. After preprocessing, the visual features of vehicle distribution, traffic speed, channel occupancy and light incident angle in the image frame are extracted.

[0007] Furthermore, the extracted visual features are tensor-encoded and the driving visual adaptation risk score is calculated in the built-in visual adaptation model, including: A structured input matrix is ​​constructed based on visual features, and position embedding and time encoding are introduced to generate a visual feature tensor with spatiotemporal continuity. The visual feature tensor is input into a multi-layer convolutional attention network structure for feature compression, and the high-response feature vector of the driving behavior attention area is extracted; Based on the preset brightness mutation sensitivity function and dynamic speed change threshold, a visual adaptation risk scoring function is constructed. The driving visual adaptation risk score corresponding to the current image frame is calculated through the feature vector and risk mapping model. The driving visual adaptation risk score reflects the probability of instantaneous visual adjustment caused by the combined effects of brightness changes, traffic density and entry angle under the current traffic conditions.

[0008] Furthermore, based on the preset brightness mutation sensitivity function and dynamic speed change threshold, a visual adaptation risk scoring function is constructed. The driving visual adaptation risk score corresponding to the current image frame is calculated through the feature vector and risk mapping model, including: Based on the brightness gradient changes in the area corresponding to the vehicle in the image frame, a brightness mutation sensitivity function is constructed. The brightness mutation sensitivity function is used to measure the spatial gradient intensity and temporal gradient change rate of the brightness in the area ahead of the vehicle, and the boundary response value of the brightness variation block is extracted through local area differential convolution. Combined with the vehicle speed change sequence, a dynamic speed change threshold function is defined to determine the vehicle's deceleration trend, speed fluctuation amplitude, and instantaneous deceleration acceleration between adjacent image frames, identifying the coupling signal between driving behavior changes and visual scene switching; A visual adaptation risk scoring function is constructed, and the brightness mutation intensity, speed change index and light incident angle are combined into a joint input feature vector. The risk mapping model and the preset scoring weight matrix are used to perform vector product mapping to calculate the driving visual adaptation risk score of the current image frame.

[0009] Furthermore, the sliding time window is combined to detect the short-term field of view mutation area and the speed reduction section, and the risk response instruction candidate set is output, including: The driving vision adaptation risk scores in consecutive image frames are input into a sliding time window of fixed length in chronological order. The gradient change value and peak duration of driving vision adaptation within the window are calculated to identify image areas with sudden brightness changes or uneven lighting. Synchronously extract the vehicle speed change sequence corresponding to each image frame within the window, and combine it with the dynamic speed change threshold function to determine whether there is deceleration behavior or speed fluctuation characteristics exceeding the set threshold, and mark it as a potential traffic interference section; The brightness mutation area is temporally aligned with the speed reduction section to construct a risk event trigger label. The trigger label is filtered according to the risk score threshold and a candidate set of risk response instructions is output.

[0010] Furthermore, the illumination adjustment module includes: Map the risk level, trigger location, and recommended response type in the risk response instruction candidate set to the preset lighting response state space, and construct a ternary state map that includes vehicle distribution density, speed trend, and light gradient change rate; In the ternary state map, the traffic flow features corresponding to the trigger position are extracted, including the number of vehicles per unit length, average speed and speed variance, as well as the light intensity gradient and brightness change direction of the corresponding area in the image frame; Based on the matching relationship between traffic flow characteristics and light gradient, the illumination adjustment rule library is called to match the corresponding lighting response level and generate the entrance section lighting illumination adjustment parameter group; The entrance section lighting illumination adjustment parameter group is used to drive the on-site lamps to perform segmented illumination adjustment, including the target illumination value, illumination gradient step, gradient duration and scope label.

[0011] Furthermore, the control command generation module includes: Extract the target vehicle's trajectory from continuous image frames, calculate the vehicle's tunnel entry behavior continuity index, and identify whether there is sudden deceleration, abnormal lane change, or trajectory interruption before entering the tunnel; Based on the vehicle speed sequence within the sliding time window and combined with historical deceleration behavior data, a short-term stop probability model is constructed to generate a corresponding vehicle stop probability score for each vehicle in the current traffic flow; Perform occlusion analysis on each lane field of view in the image frame, and calculate the blind spot occlusion score corresponding to the current frame based on the spatial projection relationship between the vehicle boundary overlap area, blind spot size, and incident light angle. The hole-entry behavior continuity index, vehicle stop probability score, and visual blind spot occlusion score are integrated to construct a signal linkage trigger factor for decision-making. The output determines whether a speed limit, guidance, or warning signal control command needs to be generated, and the corresponding signal type and scope of action are selected based on the decision result.

[0012] Furthermore, the execution results and image change sequences are collected simultaneously to extract the lighting response deviation and behavior trajectory offset as the update input of the visual adaptation model, including: Collect the execution results of the lighting illumination adjustment parameter group and signal control command, and compare them with the corresponding image frames before execution to generate the image change sequence after the lighting response; Calculate the lighting response deviation between the current light output and the preset target illuminance based on the average brightness, gradient distribution, and lane boundary clarity of the tunnel entrance area in the image change sequence; The target vehicle's trajectory before and after control is extracted, and the vehicle behavior trajectory deviation index is calculated based on the time alignment offset between trajectory point sets, trajectory shape change rate, and lane center deviation distance. The lighting response deviation and the behavioral trajectory offset index are combined into a joint feedback vector and input into the visual adaptation model as incremental update data.

[0013] Furthermore, based on the behavioral feedback results, a fine-tuning mechanism is used to update the brightness response mapping relationship and linkage strategy rules in the visual adaptation model, including: Compare the joint feedback vector composed of the lighting response deviation and the behavior trajectory offset index with the historical response state of the model to determine the dynamic convergence trend between the current control effect and the predicted deviation; When insufficient lighting response or vehicle traffic deviation exceeds the tolerance threshold in multiple consecutive time windows, the local fine-tuning mechanism of the visual adaptation model is triggered, and the mapping parameters of the affected area are selected for update; The gradient descent method is used to perform fine-grained correction on the illumination adjustment weight in the brightness response mapping relationship, so as to reduce the error between the actual lighting output and the desired visual state. At the same time, the trigger threshold, scope of action and signal type selection conditions in the signal linkage strategy rules are updated to build a new round of control strategy groups that adapt to the current traffic behavior pattern.

[0014] Furthermore, adaptive optimization of the combined control of lighting and signals in different time periods and environments is performed, including: According to the currently detected time period information and environmental status, the optimal control template is matched in the control scenario library, and a dynamic adjustment strategy is generated in combination with the latest fine-tuned model parameters; The illumination level boundary, signal type switching threshold and linkage delay rules in the matching control template are called to realize adaptive optimization of the joint control of lighting and signals in different scenarios.

[0015] The technical effects and advantages of the tunnel entrance dynamic lighting and signal control system based on visual algorithms of the present invention are as follows: This invention achieves continuous monitoring and intelligent control of driver visual adaptation in complex traffic and lighting environments at tunnel entrances by constructing a dynamic perception mechanism based on visual feature understanding and multi-strategy linkage. Based on continuous image acquisition and high-quality preprocessing, it extracts multidimensional visual features such as vehicle distribution, speed trends, channel occupancy, and light incident angle. It then generates a driver visual adaptation risk score through tensor coding and a convolutional attention mechanism. It then dynamically detects sudden visual field changes and abnormal traffic behavior using a sliding window, outputting a candidate set of risk response instructions. Based on this, the system constructs a lighting control state map, explores the matching relationship between traffic flow characteristics and lighting gradients, and generates a gradient-varying segmented illumination adjustment parameter set, improving the response accuracy and transition smoothness of lighting control.

[0016] At the same time, by integrating relevant dynamic factors such as vehicle trajectory continuity, parking probability and blind spot occlusion, a signal linkage trigger mechanism is constructed to realize the real-time issuance of speed limit, guidance and warning signals; combining the controlled image and trajectory feedback, the lighting response deviation and behavioral trajectory offset indicators are extracted, a joint feedback vector is generated and drives the local fine-tuning of the visual adaptation model, completing the online optimization and update of the brightness mapping relationship and linkage strategy rules; finally, through the control scene library and multi-environment label matching mechanism, an adaptive joint control strategy is generated across time periods and multiple conditions, which significantly improves the traffic safety, lighting accuracy and signal control intelligence level at the tunnel entrance in dynamic traffic and complex lighting environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a structural diagram of the tunnel entrance dynamic lighting and signal control system based on visual algorithms of the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0019] In order to achieve the above objectives, Figure 1 The present invention provides a schematic diagram of the structure of a tunnel entrance dynamic lighting and signal control system based on a visual algorithm. Specifically, it includes a tunnel entrance image acquisition module, a response instruction generation module, an illumination adjustment module, a control command generation module, an update input module, and a joint control module. Each module is connected by signals. The tunnel entrance image acquisition module collects a continuous sequence of image frames from the tunnel entrance area, performs defogging, contrast enhancement, and lane normalization processing, and extracts visual features from the image frames. The response instruction generation module performs tensor encoding on the extracted visual features and calculates the driving visual adaptation risk score using the built-in visual adaptation model. It then uses a sliding time window to detect short-term visual field mutation areas and speed reduction sections, and outputs a set of risk response instruction candidates. The illumination adjustment module builds a lighting control state map based on the risk response instruction candidate set, extracts the matching relationship between traffic flow characteristics and light gradients, and maps and generates a set of entrance section lighting illumination adjustment parameters; The control command generation module simultaneously constructs signal linkage trigger factors, combining the continuity of the tunnel entry behavior, the probability of the vehicle stopping, and the degree of visual blind spot obstruction to determine whether to generate speed limit, guidance, or warning signal control commands; The update input module sends the lighting illumination adjustment parameter group and signal control command to the on-site control device for execution. At the same time, it collects the execution results and image change sequence, extracts the lighting response deviation and behavior trajectory offset as the update input of the visual adaptation model; The joint control module, based on behavioral feedback results, adopts a fine-tuning mechanism to update the brightness response mapping relationship and linkage strategy rules in the visual adaptation model, and adaptively optimizes the joint control of lights and signals in different time periods and environments.

[0020] The tunnel entrance image acquisition module collects a sequence of continuous image frames from the tunnel entrance area, performs defogging, contrast enhancement, and lane normalization, and extracts visual features from the image frames, including the following: Tunnel entrance image acquisition equipment is deployed on fixed brackets above or on both sides of the tunnel entrance area. It can capture a continuous sequence of image frames covering the entire traffic lane. The image acquisition frequency is set according to the traffic flow density, with a typical frame rate of 25 frames per second. The image frames are RGB color images with a resolution of no less than 1920×1080 pixels. After the image frames are acquired, image preprocessing operations are performed in sequence, including dehazing, contrast enhancement, and lane normalization. The specific operations are as follows: An image dehazing algorithm based on a dark channel prior is employed. First, the minimum value of each pixel in each image frame across the three RGB color channels is calculated to generate a dark channel image. A local minimum filter is then used to calculate the local minimum region of this dark channel image, which is used to estimate the atmospheric light value. An image restoration model is then established to restore the reflectivity of the original image, outputting a clear image with near-true brightness. To further enhance the processing effect, a histogram equalization step is introduced to stretch the image brightness distribution and expand dark area details. This allows for clearer vehicle outlines and boundaries in image frames in low visibility or shadow conditions, thereby enhancing image clarity. To address issues with low overall brightness or uneven distribution of bright and dark areas in image frames, the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm is used for image enhancement. This method divides the image into multiple local regions and performs a histogram equalization operation on each region to enhance local contrast. During this process, a contrast limit threshold is set to prevent over-enhancement of the image, with a typical threshold range of 2.0 to 4.0. This processing step can effectively enhance the grayscale gradient of areas such as vehicle edges, lane lines, and traffic signs, thereby increasing the saliency of image edge features and enhancing the visual separation between lane edges and vehicle outlines, facilitating subsequent target extraction and behavior discrimination. Lane normalization is performed. To maintain image analysis consistency under different camera installation angles, tunnel structures, and viewing angle variations, image perspective correction is combined with a lane template matching algorithm to perform lane normalization on the image frames. The specific steps are: extracting the perspective transformation matrix from fixed road marking points in the image; using the perspective transformation matrix to perform a homography transformation on the original image to eliminate tilt and distortion in the image, so that the lane lines are in a standard parallel state in the transformed image; matching and comparing the corrected image with the preset standard lane template to ensure that the lane area is in a unified spatial coordinate reference system, thereby achieving processing uniformity and algorithm generalization capabilities in different scenarios; After completing the above image preprocessing, multi-dimensional visual features for subsequent risk scoring are extracted from each frame of the image. The extracted features include but are not limited to: Use the YOLOv5 object detection network to identify and annotate vehicle targets in the image frame, extracting the location coordinates (center point coordinates and bounding box size) and lane number of each vehicle; Based on a multi-frame target tracking algorithm (such as DeepSORT), the vehicle displacement and time interval in consecutive image frames are calculated to obtain the vehicle velocity vector, and the velocity change between the current frame and the previous frame is recorded; Divide the lane into fixed area grids and count the percentage of grids blocked by vehicles in each lane to represent the current traffic density. Combining the vehicle shadow direction and the bright boundary gradient direction in the image frame, a projection geometry method is used to calculate the projection angle of the main direction of the light source in the lane coordinate system to reflect the impact of the current external tunnel lighting on the driver's field of view.

[0021] In summary, through continuous image frame acquisition and multi-dimensional image preprocessing (including dehazing, contrast enhancement, and lane normalization), it is possible to extract clear, structurally consistent image inputs under different weather conditions, lighting conditions, and camera angles, providing a reliable data foundation for subsequent visual feature analysis and significantly improving image readability.

[0022] The response instruction generation module performs tensor encoding on the extracted visual features and calculates the driving visual adaptation risk score using the built-in visual adaptation model. It then uses a sliding time window to detect short-term visual field mutation areas and speed reduction sections, and outputs a set of risk response instruction candidates, as follows: The system receives visual feature data extracted by the tunnel entrance image acquisition module. These visual features include multi-dimensional information such as vehicle distribution, traffic speed, channel occupancy, and light incident angle. First, these visual features are temporally encoded at a fixed frame interval to construct a structured input matrix. Each row in the structured input matrix corresponds to a visual feature type, and each column corresponds to a spatial location unit within the image frame after gridding. To maintain the integrity of the spatial information, a position embedding vector is introduced for each location unit in the input matrix. This vector encodes the location of the unit in two-dimensional space within the image frame. Furthermore, to characterize the temporal evolution of visual features in a multi-frame image sequence, temporal encoding is introduced. Each frame is assigned a time step number, and the temporal dimension is represented using sine and cosine functions or a learnable embedding vector. Through the dual injection of spatial position embedding and temporal encoding, a visual feature tensor with a clear spatiotemporal sequence structure is constructed. This visual feature tensor is a five-dimensional structure corresponding to the batch dimension, the number of time frames, the number of feature channels, the spatial height, and the spatial width, and has strong temporal representation capabilities for traffic status.

[0023] The generated visual feature tensor is fed into a multi-layer convolutional attention network within the visual adaptation model for feature compression and focused representation. The network structure comprises multiple 3D convolutional layers that simultaneously extract temporal variations and spatial pattern features. The channel attention module scores and weights the responsiveness of different feature channels to improve the responsiveness of channels sensitive to driving behavior. The spatial attention module analyzes the responsiveness differences between spatial locations in the tensor to locate regions in the current frame that significantly influence driving decisions (such as the edge of the visual field, the critical area at the tunnel entrance, and the leading edge of the blind spot). The extracted features of these highly responsive regions are aggregated into a unified feature vector, forming a high-dimensional behavioral representation of the driving state. This feature vector is then fed into the visual adaptation risk score function, which calculates the risk score function to generate a risk score that reflects the driver's degree of visual stress or discomfort in the current traffic state. This risk score is a continuous floating-point number between 0 and 1. Higher values ​​indicate a greater impact on the driver's visual field adaptability due to factors such as sudden brightness changes, increased traffic density, or unreasonable tunnel entry angles. Lower values ​​indicate a stable visual environment and a favorable traffic state.

[0024] The construction of the visual adaptation risk scoring function is based on three key input features: the intensity of the brightness mutation in the area ahead of the vehicle in the image frame, the vehicle speed change index, and the spatial projection information of the incident light angle in the image frame; The brightness mutation sensitivity function is used to measure the instantaneous impact of brightness changes on the driver's vision; A sensitivity function for brightness changes is constructed and calculated as follows: First, the brightness distribution of a specified area in front of the target vehicle (e.g., 5 to 15 meters in front of the vehicle) is extracted from the image frame and divided into several subgrids. The average brightness values ​​of each subgrid between adjacent frames are then differentiated to calculate the spatial and temporal brightness gradients. The spatial gradient represents the degree of brightness change between adjacent regions within a given frame, while the temporal gradient represents the rate of change in brightness within that region between frames. Subsequently, a local differential convolution operation (i.e., a convolution kernel with a brightness gradient calculated within a 3×3 pixel range) is performed to extract regions with sharp brightness changes at their boundaries. The average gradient value of all these regions is used as the brightness change intensity indicator. For example, if the brightness of the area in front of the vehicle rapidly increases from 80 units to 200 units at a given moment in time, with a time interval of 0.2 seconds, the average temporal gradient is (200–80) / 0.2 = 600 units / second, which serves as an input to the sensitivity function. In order to characterize the speed fluctuation of the vehicle caused by sudden environmental changes, a dynamic speed change threshold function is defined, and its calculation method is as follows: The speed of the same target vehicle in consecutive image frames is recorded as a time series, and a sliding time window is used to monitor speed changes. Within each time window, the maximum deceleration, speed change variance, and speed direction variation are calculated. If the speed drop value of the current frame is greater than 1.5 times the historical mean drop, or the speed fluctuation variance exceeds the preset warning threshold, it can be determined that abnormal deceleration behavior is currently occurring. The result of this indicator is output in the form of a numerical score, for example, a score range of 0 to 1 is set to represent a trend from no significant deceleration to strong deceleration or stopping.

[0025] The output process of the risk response instruction candidate set is as follows: Combining the brightness mutation intensity, speed change score, and incident light angle, a joint input feature vector is constructed and fed into the risk score mapping model. This model uses a predefined score weight matrix to assign different weights to different feature components. For example, a weight of 0.5 is assigned to brightness mutation intensity, 0.3 to speed fluctuation, and 0.2 to incident light angle. The final driving visual adaptation risk score is calculated by weighted multiplication and summation of these components. For example, if the brightness score is 0.9, the speed score is 0.6, and the incident light angle score is 0.4, the score is: 0.9 × 0.5 + 0.6 × 0.3 + 0.4 × 0.2 = 0.45 + 0.18 + 0.08 = 0.71. This risk score is output as a floating-point value, indicating the degree to which the visual environment corresponding to the current image frame affects the driver's adaptability. Higher scores indicate stronger visual mutations and more unstable driving behavior, triggering a coordinated control response for lights and signals.

[0026] The driving vision adaptation risk score values ​​calculated from continuous image frames are input into a sliding time window of fixed length according to the image time sequence. The length of the time window can be set according to the actual traffic speed and system response time. It is generally recommended to set it to cover an image frame sequence of 1 to 3 seconds. For example, at a sampling frequency of 25 frames per second, the window length can be set to 75 frames.

[0027] Within each sliding window, the system calculates the gradient change value and local peak duration of the score curve. The gradient change value refers to the rate of change in the score between consecutive frames, i.e., the magnitude of the score increase or decrease, which is used to identify the time point of a sudden score change. The peak duration represents the number of frames during which the score remains in the high-risk range (e.g., greater than 0.7), which is used to determine the duration of illumination interference. When a sudden increase in the score is detected within a window (e.g., from 0.3 to 0.85 within a short period of time and lasting for more than 10 frames), the system deems the current image frame to have a sudden change in field of view brightness or a rapid change in uneven illumination, and locates it as a potential area of ​​visual impact. In parallel with the visual scoring, the system also extracts speed change data for each vehicle from the image frames, constructing a vehicle speed change sequence. Each entry in the sequence represents the speed difference between consecutive frames for the target vehicle. The system then applies the dynamic speed change threshold function defined in the previous embodiment to evaluate the speed change sequence and determine whether there is any deceleration that exceeds the normal range. For example, if the speed of the same target vehicle drops from 40 km / h to 10 km / h in a short period of time within the sliding window, or the speed fluctuates by more than 2 m / s for three consecutive frames, the system will determine that the vehicle is currently in a state of potential abnormal deceleration. The period that meets this condition is marked as a potential traffic interference section, which will be used for comparison with the field of view mutation area.

[0028] The system then performs temporal registration of the brightness change regions and the speed reduction sections. Specifically, the timestamps (i.e., image frame numbers) of the two events are compared to determine whether they occur within the same sliding window period and whether there is overlap or proximity. If the two events occur simultaneously or nearly simultaneously (e.g., within a timeframe of no more than five frames), a joint event label, called a risk event trigger label, is constructed. The system further filters these trigger labels based on a risk score threshold. The default threshold is set to 0.7, indicating that only those with a score greater than this threshold are considered visually impactful. Eligible trigger labels are output as a set of risk response instruction candidates. Each candidate includes the event's location, duration, recommended lighting enhancement level, and whether a simultaneous signal instruction (such as a speed limit or guide signal) is recommended. This set of risk response instruction candidates is subsequently used for decision matching and response execution in the illumination adjustment module and the joint control module.

[0029] In summary, by tensoring visual features and constructing a visual adaptive risk scoring model, combined with a sliding time window to detect risk mutation trends, rapid identification of sudden visual field changes and traffic anomalies can be achieved, effectively supporting the system to judge potential risk areas in advance and output candidate response strategies, thereby enhancing the system's foresight and early warning capabilities.

[0030] The illumination adjustment module builds a lighting control state map based on the risk response instruction candidate set, extracts the matching relationship between traffic flow characteristics and light gradients, and maps and generates the entrance section lighting illumination adjustment parameter group, as follows: Based on the candidate set of risk response instructions output by the preceding module, the risk level (score), trigger location (spatial coordinates), and recommended response type (e.g., lighting enhancement, signal guidance) in each instruction are parsed. This information is then mapped as a ternary mapping vector to a preset lighting response state space. This lighting response state space is a set of rules generated by historical control data and manual experience, used to characterize the correspondence between traffic flow conditions, lighting changes, and required lighting levels. In this lighting response state space, the three core features of the state map are defined as: vehicle distribution density, speed trend, and lighting gradient change rate. These three describe the congestion of the traffic space, vehicle behavior trends, and brightness fluctuations in the tunnel section, respectively. During the state map construction process, actual traffic flow characteristic parameters are extracted from the image frame area corresponding to the trigger position. Specifically, vehicle distribution density refers to the number of vehicles per unit space length. The system divides the tunnel entrance section into equal-length grids (for example, 10 meters per section) and counts the number of areas blocked by vehicles in each section as a vehicle density indicator. The speed trend is represented by the average and changing trend of the speed of each vehicle in the sliding time window. The average speed and speed variance of vehicles in consecutive frames in the target lane are calculated to determine whether there is a trend of deceleration, aggregation, or congestion. The illumination gradient change rate represents the magnitude of the change in regional brightness distribution with position in the image frame. The rate of change of brightness difference in the spatial dimension is used to calculate the degree and directionality of illumination mutation combined with the average grayscale difference between the current frame and the previous frame. Based on the above three types of traffic characteristics, the most suitable lighting response level is matched in the illumination adjustment rule library. The illumination adjustment rule library is a preset multi-level response strategy set that contains the lighting adjustment strategies required for different combinations of vehicle density, speed and light mutation. For example, when the vehicle density is high, the speed drops significantly and the light gradient mutation is drastic, the corresponding lighting response level is enhanced mode 3, that is, fast brightness, high illumination, and long gradient; when the vehicle density is low, the speed is stable but the light mutation is slight, only gradual brightness mode 1 may be required, that is, medium brightness and short transition. Based on the matching results, the system generates the entrance section lighting illumination adjustment parameter group, which includes: Target illuminance value: This is the lighting intensity expected to be achieved in the current section, in lux. It is set dynamically based on the tunnel design standard and risk level. For example, the standard value is 400 lux, which can be increased to 600 lux during high-risk periods. Illumination gradient step size: This is the increment of each change in light illumination adjustment to avoid sudden changes that may cause new visual impacts. For example, it can be set to increase by 20 Lux per frame. Gradual duration: controls the time required to smoothly transition from the current illuminance to the target illuminance, in seconds, and is usually dynamically adjusted within the range of 1 second to 5 seconds; Scope label: used to indicate which segmented lamps need to participate in the adjustment. For example, the three segments of lighting units within the range of 0 to 30 meters at the tunnel entrance need to execute this parameter group command at the same time.

[0031] The illumination adjustment parameter group is sent to the on-site lighting controller through the system control interface, driving the lamps in the corresponding area to perform lighting adjustments according to the set illumination value, change rhythm and dimming range, forming an intelligent lighting control strategy that is synchronized with traffic dynamics and visual changes.

[0032] The illumination adjustment module constructs a lighting control state map and generates a segmented illumination adjustment parameter group based on the relationship between traffic flow and light gradient. This can achieve dynamic graded control of the light intensity at the tunnel entrance section, effectively alleviating the visual impact at the junction of light and dark, achieving a smoother brightness transition, and improving the driver's adaptability to entering the tunnel and traffic safety.

[0033] The control command generation module simultaneously constructs a signal linkage trigger factor. It combines the continuity of the tunnel entry behavior, the probability of the vehicle stopping, and the degree of visual blind spot obstruction to determine whether to generate a speed limit, guidance, or warning signal control command. The details are as follows: Based on the vehicle target position coordinates extracted from consecutive image frames, each vehicle is tracked across frames to form a vehicle trajectory. The trajectory is a chronological sequence of vehicle center point coordinates, and can also be expanded to record information such as vehicle speed, lane number, and bounding box. From this, a tunnel entry behavior continuity index is calculated. This index is used to evaluate the smoothness of the target vehicle's movement from the tunnel approach section to the tunnel entrance. It mainly includes the following three behavioral pattern recognitions: Sudden deceleration: Within 20 meters before entering a tunnel, if the vehicle speed drops by more than a set threshold (e.g., from 50 km / h to 20 km / h) within three frames, it is considered discontinuous deceleration. Abnormal lane change behavior: If the vehicle's trajectory deviates laterally within a short period of time and the lane change behavior does not comply with the target lane rules (such as crossing the line or deviating from the center line by more than 2 meters), it will be marked as an abnormal lane change; Trajectory interruption: If the target vehicle has inter-frame missing in continuous images (such as not being detected for more than 3 frames) or the positioning jump exceeds the spatial step allowed by the normal speed, it is considered a trajectory interruption.

[0034] The above three types of behaviors are uniformly mapped into a hole entry behavior continuity score, ranging from 0 to 1. The lower the value, the worse the continuity and the more likely it is to trigger signal control requirements. Based on the vehicle speed sequence and combined with the deceleration behavior patterns in historical traffic samples, a short-term parking probability model is constructed. The model input includes the speed mean, speed variance and deceleration distribution of each vehicle in the sliding time window; historical data can be collected through early deployment or generated by the simulation system. The model uses a probability estimation method based on K-nearest neighbor discrimination or Bayesian statistics to output a probability score for each vehicle's current behavior that it will stop or approach zero speed in the next 1 second. For example, when the vehicle's instantaneous speed change rate exceeds three meters per second squared, that is, the speed of the vehicle increases or decreases by more than three meters per second in one second, and there are two temporary near-zero speed behaviors in the sliding window, its stop probability score may be assigned a value of 0.75. The system sets a stop probability score threshold (such as 0.6), and those exceeding it are marked as high-risk vehicles; Identify the occlusion risk of the blind spot in the current frame, perform occlusion analysis on the image area of ​​each lane, and calculate the blind spot occlusion score. The blind spot occlusion score is composed of the following factors: By detecting the area ratio of overlapping areas of multiple vehicle bounding boxes, the system determines whether occlusion overlap occurs. Based on the vehicle density and camera field of view, the system dynamically estimates the area that the camera cannot effectively capture. The system then combines the incident light angle with the angle between the lanes to assess the probability that oblique sunlight will cause shadows to block the camera's field of view. The smaller the incident angle and the longer the projected shadow, the higher the occlusion score. The above three scoring factors are combined and the final visual blind spot occlusion score is calculated according to the set weights (for example, boundary overlap 0.4, blind spot size 0.3, light occlusion 0.3, weighted to 0.4×0.5+0.3×0.3+0.3×0.2=0.37). The higher the score, the greater the impact of the current visual blind spot. Ultimately, the system inputs the hole-entry continuity score, the vehicle stop probability score, and the visual blind spot obstruction score into the signal control decision logic to construct a signal linkage trigger factor. This signal linkage trigger factor is a ternary structure vector representing the three risk intensities of the current traffic conditions. It is equipped with response weights and multi-layered judgment rules. When any two of the three scores exceed a set risk threshold (e.g., both above 0.7), the system outputs a signal control trigger flag and selects the corresponding signal response based on a recommended signal type mapping table. For example, if continuity is low and obstruction is high, a speed limit prompt is issued; if the probability of stopping is high and continuity is medium, a guide signal is issued; and if all three are high, a warning signal is issued. Furthermore, based on the vehicle's position and the extent of the obstruction area, the signal control command's area of ​​effect is labeled (e.g., the left lane of the entrance section, the 10-30 meter section of the right lane, etc.), resulting in a complete output of structured signal control instructions.

[0035] The control command generation module constructs signal linkage trigger factors and generates precise control instructions through a comprehensive assessment of vehicle behavior continuity, parking probability, and the degree of visual blind spot obstruction. It can trigger speed limit, guidance, or warning signals before the vehicle is about to enter a risky state, thereby improving the targeted release of signals and the timeliness of linkage, and reducing accident risks.

[0036] The update input module sends the lighting illumination adjustment parameter group and signal control command to the on-site control device for execution. At the same time, it collects the execution results and image change sequence, extracts the lighting response deviation and behavior trajectory offset as the update input of the visual adaptation model, as follows: The lighting illumination adjustment parameter group and signal control command generated by the illumination adjustment module and the control command generation module are sent to the on-site control equipment, specifically including: transmitting the parameter group to the lighting control driver and signal control terminal through a wired or wireless communication link, controlling them to respectively perform light brightness adjustment, gradient step adjustment, range start and stop, and signal prompt graphic display, color switching and guide sign control; After the control command is executed, the image feedback acquisition program is started synchronously to record the continuous image frame sequence after the command execution. These image frames are then matched one by one with the image frames before the control execution in chronological order to generate a sequence of image changes after the lighting response. The specific matching method is to accurately align the timestamps and use the image similarity matching algorithm to correct inter-frame drift when necessary. Based on this image change sequence, the system extracts three key indicators for the tunnel entrance area to calculate the lighting response deviation: Calculate the average brightness, which is the grayscale mean of all pixels within the designated area of ​​the tunnel entrance in the current image frame, to reflect the overall lighting level. Calculate the brightness gradient distribution, which evaluates the smoothness of lighting transitions by calculating the spatial first-order derivative of the image grayscale value to prevent light bounces or localized overbrightness. Calculate the lane boundary clarity index: Based on an edge detection algorithm, this index evaluates the edge strength and continuity of lane boundary lines. This index indirectly determines whether the lighting is conducive to lane structure recognition. The above indicators are calculated by subtracting the target illumination value, expected gradient change rate, and other control target parameters to obtain the lighting response deviation of the current frame. For example, if the target brightness is set to 600 Lux and the average brightness of the captured image is only 500 Lux, the response deviation is -100 Lux. If the gradient change rate deviates from the expected value by more than 20%, this will also be recorded as a deviation. At the same time, the trajectory of each target vehicle is extracted from the image frames before and after the control execution, and the vehicle's trajectory point sequence in the image coordinate system is obtained through a multi-frame tracking algorithm. The time alignment offset (i.e., the difference in the number of frames of trajectory delay or advance), trajectory morphology change rate (i.e., the mean square error of changes in trajectory curvature, broken line angle, etc.), and lane center deviation distance (i.e., the average vertical distance of the vehicle trajectory relative to the ideal lane centerline) of the trajectory are calculated in the same time period before and after the control. These indicators together constitute the vehicle behavior trajectory deviation index, which is used to measure whether the lighting and signal adjustments have an actual guiding or interference effect on driving behavior. The lighting response deviation and behavioral trajectory deviation indicators are integrated into a set of joint feedback vectors, which represent the environmental visual changes and behavioral response deviations after each control response in a standardized manner and are input into the model training interface as incremental update data of the visual adaptation model.

[0037] The update input module compares and analyzes the image changes and vehicle behavior trajectory before and after the execution of the control command, extracts the lighting response deviation and behavior trajectory offset indicators, and constructs a joint feedback vector to provide quantitative support for model updating, realize closed-loop feedback adjustment, and improve the learning efficiency and update accuracy of the visual adaptation model.

[0038] The joint control module, based on behavioral feedback, uses a fine-tuning mechanism to update the brightness response mapping relationship and linkage strategy rules in the visual adaptation model, and adaptively optimizes the joint control of lighting and signals in different time periods and environments. Specific contents include: After each control cycle, the lighting response deviation and the behavioral trajectory deviation index form a joint feedback vector. The former describes the difference between the target illuminance value and the actual luminance output, while the latter describes the temporal and spatial offset between the vehicle's path and the ideal lane path. The system compares this joint feedback vector with the historical model response state, including: Match historical execution data under the same environmental label or control scenario to analyze the convergence or divergence trend between the current predicted lighting and vehicle behavior status and historical results; If the system detects any of the following abnormal conditions within multiple consecutive time windows (for example, three consecutive sliding windows, each with 30 frames): first, insufficient lighting response, i.e., the actual average brightness continuously falls below the target illuminance by a certain range (e.g., the deviation exceeds 50 Lux); second, vehicle deviation, i.e., the target vehicle's trajectory center deviates from the ideal lane center by more than a set tolerance (e.g., greater than 0.8 meters), then the local fine-tuning mechanism of the visual adaptation model is triggered; After the fine-tuning mechanism is activated, the model sub-parameter set corresponding to the affected area is first located. For example, when the current triggering area is the 0 to 20 meter section of the tunnel entrance and the risk score is high, the system will extract the brightness response mapping parameters that match the label and risk level of the area, including the illuminance output weight associated with the area, the brightness increment step factor, the gradient time adjustment factor, etc. Subsequently, the system uses the gradient descent method to fine-tune and optimize the sub-parameter set: Specifically, the system uses the error value between the expected visual state and the actual response state as the loss target, and updates the illuminance adjustment weight parameter by gradually iteratively reducing the error. For example, when the target illuminance is 600Lux and the actual illuminance is 520Lux, the system will look for the optimal path in the parameter weight matrix to increase the output to close to 600Lux, and appropriately increase the basic illuminance parameters and adjustment rate of the area; Further structural updates can be made to the "Signal Linkage Strategy Rules" simultaneously, including: Trigger threshold updates, such as those for vehicle stop probability scores or blind spot obstruction scores, will be appropriately increased or decreased based on feedback. Range updates will adjust the lane or distance range of signal effects based on actual feedback regarding vehicle offset ranges. Signal type selection conditions will be updated, meaning that if the current template repeatedly triggers a certain type of signal but fails to effectively change vehicle behavior, the system will consider switching to a stronger warning signal or extending the signal duration. Based on the above fine-tuning results, a new set of control strategies is constructed and added to the control strategy history library. This fine-tuning process is a local, lightweight adjustment that does not change the overall model structure and only optimizes within the parameter space to ensure online operation efficiency and system stability.

[0039] Adaptive optimization of combined lighting and signal control in different time periods and environments, including: During the initial deployment phase, a variety of typical control scenarios were extracted through long-term operational data and expert strategies, and a control scenario library was constructed. Each scenario in this library is identified by a set of environmental tags, including but not limited to: time tags (e.g., morning rush hour, nighttime, dusk), weather tags (e.g., strong sunlight on a sunny day, rainy and foggy days), traffic tags (e.g., dense traffic, intermittent traffic), and illumination state tags (e.g., light-dark junction sections). Each control scenario is associated with a set of control templates, which contain preset recommended illumination ranges, signal linkage strategies, response delay adjustment parameters, and parameter correction coefficients. Through real-time data collected by sensors and cameras, the system detects the time period of the current scene (such as whether it is dusk in combination with the system clock and the local sunlight schedule), lighting conditions (such as whether the image brightness mean and gradient value are combined to assess whether there is strong backlight), traffic information (such as whether the vehicle unit area density and speed variance determine whether it is a high-density flow) and weather conditions (such as obtained from the external environment data interface), forming a set of current environment status labels.

[0040] During the scene matching process, the system compares the similarity between the current tag set and the tag features of each template in the control scene library. It can use weight judgment, tag voting or priority matching strategy to match. For example, if the current tag meets the three tags of night, low speed and strong occlusion at the same time, the system will select the night low speed and high occlusion template with the closest tag weight as the optimal control template. The matched control template contains multiple key parameters, including: Illumination level boundaries: the minimum and maximum illuminance values ​​for each risk level (e.g., 400–550 Lux for medium risk and 600–750 Lux for high risk); Signal type switching threshold: A combined scoring threshold used to determine whether to switch from a speed limit signal to a warning signal (e.g., behavior continuity score <0.3 and blind spot score >0.7); Linkage delay rule: used to determine the time offset between lighting changes and signal switching to prevent synchronization shock (for example, the lighting gradient changes 2 seconds in advance and the signal starts 1 second later).

[0041] The above parameters are combined with the current fine-tuned visual adaptation model to generate a set of dynamic adjustment strategies. This strategy is continuously refreshed, called, and issued for execution during the control cycle. The specific implementation includes: the lamps adjust the light intensity within the matching illumination level boundaries, the controller dynamically determines whether to issue speed limit, guidance or warning instructions based on the signal switching threshold, and coordinates the rhythm of lighting and signal control according to the linkage delay rules.

[0042] The joint control module fine-tunes model parameters based on behavioral feedback results, and combines it with the control scenario library to achieve adaptive scheduling of lighting and signal strategies, effectively improving the system's generalization capabilities and control stability in different time periods, different traffic densities, and different lighting environments, ensuring that the strategy response is not rigid and the control does not fail during the long-term operation of the system.

[0043] The above embodiments may be implemented in whole or in part through software, hardware, firmware or any other combination. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0044] Those skilled in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0045] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0046] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0047] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A tunnel entrance dynamic lighting and signal control system based on a visual algorithm, characterized by: It includes a tunnel entrance image acquisition module, a response instruction generation module, an illumination adjustment module, a control command generation module, an update input module, and a joint control module, and each module is connected by signals; The tunnel entrance image acquisition module collects a continuous sequence of image frames from the tunnel entrance area, performs defogging, contrast enhancement, and lane normalization processing, and extracts visual features from the image frames. The response instruction generation module performs tensor encoding on the extracted visual features and calculates the driving visual adaptation risk score using the built-in visual adaptation model. It then uses a sliding time window to detect short-term visual field mutation areas and speed reduction sections, and outputs a set of risk response instruction candidates. The illumination adjustment module builds a lighting control state map based on the risk response instruction candidate set, extracts the matching relationship between traffic flow characteristics and light gradients, and maps and generates a set of entrance section lighting illumination adjustment parameters; The control command generation module simultaneously constructs signal linkage trigger factors, combining the continuity of the tunnel entry behavior, the probability of the vehicle stopping, and the degree of visual blind spot obstruction to determine whether to generate speed limit, guidance, or warning signal control commands; The update input module sends the lighting illumination adjustment parameter group and signal control command to the on-site control device for execution. At the same time, it collects the execution results and image change sequence, extracts the lighting response deviation and behavior trajectory offset as the update input of the visual adaptation model; The joint control module, based on behavioral feedback results, adopts a fine-tuning mechanism to update the brightness response mapping relationship and linkage strategy rules in the visual adaptation model, and adaptively optimizes the joint control of lights and signals in different time periods and environments.

2. The tunnel entrance dynamic lighting and signal control system based on a visual algorithm according to claim 1 is characterized by: Collect a sequence of continuous image frames from the tunnel entrance area, perform defogging, contrast enhancement, and lane normalization, and extract visual features from the image frames, including: Dehazing the image frame is performed using dark channel prior and histogram equalization algorithm; Performing contrast enhancement on image frames using contrast-limited adaptive histogram equalization method; Based on image perspective correction and lane template matching algorithms, lane normalization is performed on image frames to convert lane structures from different perspectives into a unified spatial reference system. After preprocessing, the visual features of vehicle distribution, traffic speed, channel occupancy and light incident angle in the image frame are extracted.

3. The tunnel entrance dynamic lighting and signal control system based on a visual algorithm according to claim 2 is characterized by: The extracted visual features are tensor-encoded and the driving visual adaptation risk score is calculated in the built-in visual adaptation model, including: A structured input matrix is ​​constructed based on visual features, and position embedding and time encoding are introduced to generate a visual feature tensor with spatiotemporal continuity. The visual feature tensor is input into the multi-layer convolutional attention network structure of the visual adaptation model for feature compression, and the high-response feature vector of the driving behavior attention area is extracted; Based on the preset brightness mutation sensitivity function and dynamic speed change threshold, a visual adaptation risk scoring function is constructed through the visual adaptation model. The driving visual adaptation risk score corresponding to the current image frame is calculated through the feature vector and risk mapping model. The driving visual adaptation risk score reflects the probability of instantaneous visual adjustment caused by the combined effects of brightness changes, traffic density and entry angle under the current traffic conditions.

4. The tunnel entrance dynamic lighting and signal control system based on a visual algorithm according to claim 3 is characterized by: Based on the preset brightness mutation sensitivity function and dynamic speed change threshold, a visual adaptation risk scoring function is constructed. The driving visual adaptation risk score corresponding to the current image frame is calculated through the feature vector and risk mapping model, including: Based on the brightness gradient changes in the area corresponding to the vehicle in the image frame, a brightness mutation sensitivity function is constructed. The brightness mutation sensitivity function is used to measure the spatial gradient intensity and temporal gradient change rate of the brightness in the area ahead of the vehicle, and the boundary response value of the brightness variation block is extracted through local area differential convolution. Combined with the vehicle speed change sequence, a dynamic speed change threshold function is defined to determine the vehicle's deceleration trend, speed fluctuation amplitude, and instantaneous deceleration acceleration between adjacent image frames, identifying the coupling signal between driving behavior changes and visual scene switching; A visual adaptation risk scoring function is constructed, and the brightness mutation intensity, speed change index and light incident angle are combined into a joint input feature vector. The risk mapping model and the preset scoring weight matrix are used to perform vector product mapping to calculate the driving visual adaptation risk score of the current image frame.

5. The tunnel entrance dynamic lighting and signal control system based on a visual algorithm according to claim 4 is characterized by: Combined with the sliding time window to detect short-term visual field mutation areas and speed reduction sections, the risk response instruction candidate set is output, including: The driving vision adaptation risk scores in consecutive image frames are input into a sliding time window of fixed length in chronological order. The gradient change value and peak duration of driving vision adaptation within the window are calculated to identify image areas with sudden brightness changes or uneven lighting. Synchronously extract the vehicle speed change sequence corresponding to each image frame within the window, and combine it with the dynamic speed change threshold function to determine whether there is deceleration behavior or speed fluctuation characteristics exceeding the set threshold, and mark it as a potential traffic interference section; The brightness mutation area is temporally aligned with the speed reduction section to construct a risk event trigger label. The trigger label is filtered according to the risk score threshold and a candidate set of risk response instructions is output.

6. The tunnel entrance dynamic lighting and signal control system based on a visual algorithm according to claim 5 is characterized by: The illumination adjustment module includes: Map the risk level, trigger location, and recommended response type in the risk response instruction candidate set to the preset lighting response state space, and construct a ternary state map that includes vehicle distribution density, speed trend, and light gradient change rate; In the ternary state map, the traffic flow features corresponding to the trigger position are extracted, including the number of vehicles per unit length, average speed and speed variance, as well as the light intensity gradient and brightness change direction of the corresponding area in the image frame; Based on the matching relationship between traffic flow characteristics and light gradient, the illumination adjustment rule library is called to match the corresponding lighting response level and generate the entrance section lighting illumination adjustment parameter group; The entrance section lighting illumination adjustment parameter group is used to drive the on-site lamps to perform segmented illumination adjustment, including the target illumination value, illumination gradient step, gradient duration and scope label.

7. The tunnel entrance dynamic lighting and signal control system based on a visual algorithm according to claim 6 is characterized by: The control command generation module includes: Extract the target vehicle's trajectory from continuous image frames, calculate the vehicle's tunnel entry behavior continuity index, and identify whether there is sudden deceleration, abnormal lane change, or trajectory interruption before entering the tunnel; Based on the vehicle speed sequence within the sliding time window and combined with historical deceleration behavior data, a short-term stop probability model is constructed to generate a corresponding vehicle stop probability score for each vehicle in the current traffic flow; Perform occlusion analysis on each lane field of view in the image frame, and calculate the blind spot occlusion score corresponding to the current frame based on the spatial projection relationship between the vehicle boundary overlap area, blind spot size, and incident light angle. The hole-entry behavior continuity index, vehicle stop probability score, and visual blind spot occlusion score are integrated to construct a signal linkage trigger factor for decision-making. The output determines whether a speed limit, guidance, or warning signal control command needs to be generated, and the corresponding signal type and scope of action are selected based on the decision result.

8. The tunnel entrance dynamic lighting and signal control system based on a visual algorithm according to claim 7 is characterized by: The execution results and image change sequences are collected simultaneously, and the lighting response deviation and behavior trajectory offset are extracted as update inputs for the visual adaptation model, including: Collect the execution results of the lighting illumination adjustment parameter group and signal control command, and compare them with the corresponding image frames before execution to generate the image change sequence after the lighting response; Calculate the lighting response deviation between the current light output and the preset target illuminance based on the average brightness, gradient distribution, and lane boundary clarity of the tunnel entrance area in the image change sequence; The target vehicle's trajectory before and after control is extracted, and the vehicle behavior trajectory deviation index is calculated based on the time alignment offset between trajectory point sets, trajectory shape change rate, and lane center deviation distance. The lighting response deviation and the behavioral trajectory offset index are combined into a joint feedback vector and input into the visual adaptation model as incremental update data.

9. The tunnel entrance dynamic lighting and signal control system based on a visual algorithm according to claim 8, characterized in that: Based on the behavioral feedback results, a fine-tuning mechanism is used to update the brightness response mapping relationship and linkage strategy rules in the visual adaptation model, including: Compare the joint feedback vector composed of the lighting response deviation and the behavior trajectory offset index with the historical response state of the model to determine the dynamic convergence trend between the current control effect and the predicted deviation; When insufficient lighting response or vehicle traffic deviation exceeds the tolerance threshold in multiple consecutive time windows, the local fine-tuning mechanism of the visual adaptation model is triggered, and the mapping parameters of the affected area are selected for update; The gradient descent method is used to perform fine-grained correction on the illumination adjustment weight in the brightness response mapping relationship, so as to reduce the error between the actual lighting output and the desired visual state. At the same time, the trigger threshold, scope of action and signal type selection conditions in the signal linkage strategy rules are updated to build a new round of control strategy groups that adapt to the current traffic behavior pattern.

10. The tunnel entrance dynamic lighting and signal control system based on visual algorithms according to claim 9, characterized in that: Adaptive optimization of combined lighting and signal control in different time periods and environments, including: According to the currently detected time period information and environmental status, the optimal control template is matched in the control scenario library, and a dynamic adjustment strategy is generated in combination with the latest fine-tuned model parameters; The illumination level boundary, signal type switching threshold and linkage delay rules in the matching control template are called to realize adaptive optimization of the joint control of lighting and signals in different scenarios.

Images