Method for testing identification and response of automatic driving vehicle to signal lamp in closed site

Abstract

The invention provides a method for testing identification and response of an automatic driving vehicle to a signal lamp in a closed site, and belongs to the technical field of automatic driving testing. The method solves the problem that it is difficult to dynamically simulate a real signal lamp interaction scene and quantitatively evaluate the accuracy of a vehicle perception-decision chain in a closed test. According to the technical scheme, the method comprises the following steps: constructing a line segment topology network with directional attributes; deploying a three-dimensional adjustable signal lamp device at the node to dynamically adjust the pose; synchronously acquiring signal lamp identification result data and control response behavior data of the vehicle; calculating a time margin based on the vehicle positioning speed and the distance from the stop line, and triggering signal lamp state grading switching through comparison with a preset threshold value; a two-dimensional verification mechanism is adopted, and identification result data and control response behavior data are aligned through a time synchronization protocol to analyze causal relevance. The method is used for testing and verifying the recognition accuracy and response timeliness of an automatic driving vehicle on a signal lamp scene in a closed site.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous vehicle testing. More specifically, the present invention relates to a method for testing the recognition and response of autonomous vehicles to traffic lights in a closed area. Background Art

[0002] In closed-field autonomous driving testing, there are several limitations to verifying a vehicle's signal light recognition and response capabilities. Existing testing methods struggle to dynamically simulate real-world intersection signal light interaction scenarios, primarily due to spatial differences caused by the fixed position of signal light devices. Perspective shifts, occlusion effects, and optical distortions caused by changes in intersection geometry cannot be effectively reproduced under static deployment conditions, weakening the reliability of perception algorithm testing. Furthermore, traditional light color switching uses fixed timing logic, making it impossible to dynamically adjust the light state based on the vehicle's motion. This results in a lack of coordination between vehicle speed, acceleration, and light color switching timing, making it difficult to assess vehicle behavior in critical scenarios such as yellow light decisions and emergency braking. Existing technologies for quantitatively evaluating perception-decision links have flawed mechanisms. During testing, signal light recognition result data, such as color classification and positioning coordinates, and vehicle control response data, such as brake pressure and steering angle, are usually collected independently, lacking strict time synchronization. Because there is hardware-level clock drift between the on-board perception sensors and the control bus, which generally exceeds 10 milliseconds, and the decision window from signal light recognition to vehicle response is usually less than 500 milliseconds, timestamp deviation makes it impossible to distinguish whether the response delay is due to perception lag or decision-making errors. In addition, a quantifiable mapping model has not yet been established between perception parameters such as recognition confidence and positioning accuracy and vehicle control commands, making it difficult to locate the specific failure link in the link.

[0003] Controlling environmental interference factors also presents challenges. Closed test sites rely on natural lighting to simulate strong glare or low-light scenarios, but light intensity is constrained by time and weather, making it difficult to reliably generate specific high-intensity glare environments. Meteorological conditions such as fog concentration cannot be precisely controlled, making it difficult to quantify the degree of interference with recognition performance. These uncontrollable factors make algorithm robustness verification under extreme conditions inefficient and make it impossible to establish a standardized test baseline.

[0004] The technical root of these issues involves multi-disciplinary collaboration challenges: Dynamic traffic light position adjustment requires addressing the coupling between mechanical control accuracy and environmental interference, for example, maintaining a height tolerance within 5 centimeters. Time alignment of heterogeneous data from multiple sources must overcome the conflict between hardware clock drift and short decision windows. Vehicle positioning errors directly impact the accuracy of stop line distance calculations, with typical positioning errors reaching 1 to 3 meters. In emergency braking scenarios, a deviation of 0.5 meters can lead to misjudgment of traffic light status. These factors collectively limit the effectiveness of closed-loop test scenarios. Summary of the Invention

[0005] One objective of this invention is to provide a method for testing the recognition and response of autonomous vehicles to traffic lights within a closed environment. This method addresses the issues of non-adjustable spatial position of traffic lights and data fragmentation in the perception-decision-making process during closed testing. Traditional methods use static traffic lights that cannot simulate the changing perspectives of real intersections, and lack a mechanism to synchronize recognition results with vehicle response behavior, making it difficult to quantitatively evaluate the performance of autonomous driving systems across the entire chain, from traffic light recognition to control.

[0006] Another objective of this invention is to address the disconnect between signal light switching and vehicle motion in traditional testing. Traditional methods, with their fixed switching timing, fail to reflect the dynamic effects of vehicle speed and acceleration on light color changes. In particular, they lack quantitative testing methods for yellow light sensitivity, resulting in insufficient verification of emergency response reliability.

[0007] Another objective of this invention is to overcome the interference of positioning errors on stop line distance calculations. Vehicle GNSS positioning deviations of typically 1 to 3 meters can easily lead to misinterpretations of traffic lights during braking. Traditional threshold control lacks an error compensation mechanism, affecting the accuracy of test results.

[0008] Another purpose of this invention is to address the optical distortion caused by dynamic signal light position deviation. Mechanical adjustments resulting in suspension height deviations exceeding 5 cm or lateral offsets exceeding 10 cm can distort the light projection and reduce the effectiveness of perception system testing.

[0009] Another objective of the present invention is to eliminate the interference of multi-source data time asynchrony on causal analysis. When the vehicle's perception and control bus clock drift exceeds 10 milliseconds, while the decision window is less than 500 milliseconds, timestamp deviation can distort the performance evaluation of the perception-decision link.

[0010] Another purpose of the present invention is to prevent physical signals from interfering with simulation results during virtual testing. Traditional methods cannot isolate the actual light from the injected signal, resulting in the yellow light sensitivity test being contaminated by the physical device status.

[0011] Another purpose of the present invention is to optimize high-speed scenarios and traffic flow adaptability. Fixed acceleration thresholds are prone to triggering false interventions when vehicle speeds exceed 60 kilometers per hour, and lack a coordinated response mechanism for vehicles queuing at intersections.

[0012] Another purpose of the present invention is to compensate for the uncontrollable environmental interference. Natural lighting and fog make it difficult to stably generate standardized test scenarios, which affects the repeatability of algorithm robustness verification under extreme conditions.

[0013] Another purpose of the present invention is to address signal light recognition failures caused by ambient light variations. Fixed brightness modes in strong glare or low-light environments can cause lights to be overexposed or invisible, reducing the perception system's test coverage.

[0014] Another objective of the present invention is to prevent delayed, abnormal data from contaminating the evaluation results. When the perception-response delay exceeds 150 milliseconds or the variance is too large, traditional methods cannot automatically eliminate invalid data, affecting the credibility of the test conclusions.

[0015] To achieve these objectives and other advantages of the present invention, a method for testing the recognition and response of an autonomous driving vehicle to a signal light in a closed area is provided, comprising: S1. Digital construction of road network: Divide the roads in the closed site into line segments with directional attributes. Each line segment is defined by at least two geographic coordinate nodes. Adjacent line segments are connected to form a topological network by sharing nodes. S2. Dynamic traffic light deployment: deploying three-dimensional adjustable traffic light devices at the nodes of the topological network to adjust the suspension height, lateral offset and pitch angle in real time; S3, Multimodal Data Acquisition: Acquire signal light recognition data through vehicle-mounted perception sensors and simultaneously record vehicle control response instructions; S4, light color switching control: Based on the vehicle's positioning speed and the real-time distance from the stop line, the time required for the vehicle to reach the stop line is calculated as the time margin. The time margin is compared with a preset threshold. When the time margin meets the threshold condition, the signal light state is triggered to switch; S5. Two-dimensional verification: Synchronously collect the vehicle's traffic light recognition result data and control response behavior data. The recognition result data includes the traffic light status classification, spatial positioning coordinates and confidence level output by the perception system; the control response behavior data includes the brake pressure, throttle opening and steering angle instructions of the vehicle actuator. The two types of data are aligned through a time synchronization protocol to verify the causal relationship between the recognition result and the control response.

[0016] Preferably, in step S4, the signal light switching is triggered in a graded manner according to the time margin, specifically: S401, basic switching threshold: When the time margin is ≥ T2, it is forced to switch to the green light state; When the time margin is ≤ T1, it is forced to switch to the red light state; When T1 < time margin < T2, maintain the current light color state; where T1 and T2 meet the constraints: T2>3.0s>2.5s>T1; S402, Yellow Light Trigger Mechanism: The yellow light determination operation is performed when the following two conditions are met: i) the current time margin is less than or equal to 2.5 seconds; ii) the time margin changes from a value greater than 2.5 seconds, or it is the first time since system initialization that the time margin enters the range greater than T1 and not exceeding 2.5 seconds; The yellow light judgment operation is as follows: if the current light state is green, the green light state will be immediately switched to a flashing yellow light state, and the yellow light will automatically switch to a red light after flashing for 3 seconds; if the current light state is red or has already been yellow, the current state will remain unchanged; Each test vehicle only performs the yellow light determination operation once in a single test; while the yellow light is flashing, the execution of the S401 basic switching rule is suspended; when the vehicle passes the stop line, the yellow light determination state is reset; S403, Yellow light trigger sensitivity virtual test: During the yellow light flashing triggered by step S402, the actual signal light power is turned off and the vehicle communication interface is used to send The autonomous driving system injects a simulated signal; The analog signal states that the yellow light triggering time margin is T V , where T V ∈ {2.6s, 2.7s, 2.8s, 2.9s,3.0s}; Traverse all T V The value is set and the vehicle response behavior is recorded. After the test, the power supply to the signal light is restored; S404, Dynamic Intervention Rules: When the basic switching interval (T1, T2) is within the range and the basic switching threshold is not violated, the following interventions are performed: i) If the vehicle acceleration is less than -0.5m / s 2 , when braking suddenly, the traffic light will be forced to turn red and the red light time will be extended; ii) If acceleration > 0.5 m / s 2 , then the current status is switched to green; iii) If any state lasts for more than 10 seconds, the light will be forced to turn green.

[0017] Preferably, a positioning error compensation mechanism is added to the time margin calculation in step S4: The vehicle positioning correction vector is generated in real time by a high-precision RTK differential base station deployed on the roadside. The correction vector includes the horizontal position offset (δ X , δ Y ) and heading angle deviation δθ; the original coordinates (X V , Y V ) According to the formula (X V +δ X ·cosδθ,Y V +δ Ysinδθ) for dynamic calibration; At the same time, an elastic buffer model for the stop line position is established. In the stop line elastic buffer model, the boundary range is generated by the following formula: gray decision interval half-width W = max(0.6m, 3σ), where σ is the standard deviation of the horizontal precision factor of the current RTK differential base station positioning. When the actual distance D between the calibrated vehicle coordinates and the stop line real , satisfying |D real -D calc When |>0.5m, the virtual boundary of the stop line is automatically extended to ±W, D calc The original calculated distance from the stop line reported by the vehicle positioning system; If the vehicle position does not cross the virtual boundary, that is, in the interval [D calc -W,D calc +W], the signal light status switches according to the S4 rule, and the dual-channel data recording mechanism is started, specifically: i) First channel recording: the original calculated distance D calc The vehicle's recognition and response behavior to the signal light based on the actual distance D real The vehicle's recognition and response behavior to traffic lights based on the baseline; When the vehicle position crosses the virtual boundary, the second channel data is automatically discarded and the first channel data is output as the valid test result.

[0018] Preferably, after the dynamic traffic light is deployed in step S2, a spatial posture closed-loop verification needs to be performed, specifically: A reference target pattern is laid on the ground directly below the traffic light, and the relative position of the traffic light and the target is captured by a binocular calibration camera mounted on top of the test vehicle. When it is detected that the actual hanging height H of the signal light deviates from the set value by more than ±5cm, or the lateral offset L is greater than 10cm, the posture compensation mechanism is automatically triggered to perform posture compensation; At the same time, the optical distortion detection module is activated after each light color change. Specifically, a wide-angle monitoring camera installed on the back of the traffic light is used to capture the shape of the light spot projected by the light. If the light spot ellipticity ε>0.25, the pitch angle is automatically corrected until the roundness error ρ≤5%.

[0019] Preferably, the dual-dimensional verification in step S5 requires the establishment of a hardware-level time synchronization system, specifically: A GNSS timing module is installed in the traffic light control cabinet to generate PPS pulse signals, which are distributed to onboard sensing sensors and the vehicle control bus via a fiber optic splitter. All data streams are tagged with the IEEE 1588v2 protocol, including a UTC microsecond timestamp. To align the timing of signal light recognition data with control response instructions, a sliding time window matching algorithm is set: based on the actual signal light switching time t0, the vehicle response data peak is searched in the interval [t0-200ms, t0+500ms]. When the time difference Δt between the recognition result and the response behavior is greater than 80ms, the test data is automatically discarded and the test sequence is retriggered.

[0020] Preferably, a power isolation protocol is executed during the yellow light trigger sensitivity virtual test in step S403: i) 5 milliseconds before injecting the simulated signal, the signal light LED driver power is cut off; ii) the simulated signal is injected again after verifying that the current drops to 0A through the current sensor; iii) at the end of the test, the power supply is restored with a delay of 200 milliseconds to avoid the vehicle response window.

[0021] Preferably, in step S404 of step S4, a vehicle speed weight factor is added to the dynamic intervention rule, specifically: When the vehicle speed v>60km / h, the acceleration threshold a th Adaptive adjustment is ±0.3m / s²; For the rule of forcing the light to turn green in a continuous state, a gradient release mechanism is set: the current light color is maintained for the first 5 seconds, a yellow warning flash is started at the 6th to 8th second, and the light switches to green at the 9th second; At the same time, microwave detectors are deployed at the topological network nodes to monitor the queue length. If the number of vehicles queuing behind the stop line N ≥ 3 and the average waiting time T av >8 seconds, automatically override the acceleration judgment rule and switch to green light directly, and push the priority passage prompt code C on the roadside display priority .

[0022] Preferably, an environmental interference simulation module is added to the multimodal data acquisition in step S3: Through atomization equipment, artificial fog of controllable concentration is generated in front of the traffic light, and the ambient light intensity is simultaneously adjusted to 50,000-100,000 lux to simulate strong glare scenes, and the vehicle's signal light recognition data under extreme conditions is collected.

[0023] Preferably, a brightness adaptation mechanism is added after the spatial pose closed-loop verification, specifically: An ambient light sensor is integrated into the signal light control cabinet to monitor the ambient illumination in real time. When the ambient illumination is lower than 100 lux, the LED drive current is automatically increased to increase the light brightness to 180% of the standard value. When the ambient illumination is higher than 90,000 lux, the pulse strobe mode of the signal light is activated.

[0024] Preferably, a performance evaluation engine is deployed after the two-dimensional verification, specifically: based on the data stream after time stamp alignment, the mean and variance of the delay from signal light recognition to vehicle control response are calculated. When the delay exceeds 150ms or the variance is greater than 30ms, the vehicle control response is automatically generated. 2 When the test fails, the test is automatically marked as invalid and an alarm signal is triggered to the monitoring platform.

[0025] The present invention has at least the following beneficial effects: First, by constructing a topological network with directional attributes, dynamic mapping of intersection geometry is achieved, ensuring that signal light perspective changes are consistent with real-world scenarios. The three-dimensional adjustable signal light assembly supports real-time adjustment of suspension height, offset, and pitch angle to replicate occlusion and optical distortion effects. A two-dimensional verification mechanism aligns recognition results (light color, coordinates, and confidence) with control responses (brake pressure and steering angle) through a time synchronization protocol, accurately quantifying latency in the perception-to-decision chain and effectively locating failure points. Time margin grading thresholds dynamically link vehicle speed and distance to light color switching, realistically simulating yellow light decision-making scenarios. Virtual testing quantifies vehicle response sensitivity under different conditions by turning off the physical light and injecting simulated signals with a time margin of 2.6 to 3.0 seconds. Dynamic intervention rules force light color switching based on acceleration direction to verify control robustness during sudden braking or acceleration through the intersection.

[0026] Second, the horizontal position offset and heading angle deviation generated by the RTK differential base station compensate for vehicle positioning errors, enabling dynamic calibration of the original coordinates. The stopping line elastic buffer model establishes a virtual boundary using the larger of 0.6 meters and 3σ. When the deviation between the calibrated distance and the actual distance exceeds 0.5 meters, the tolerance range is expanded to prevent false triggering caused by positioning noise. Dual-channel data logging stores both the original calculated distance and the actual distance test results, ensuring the output data is consistent with the vehicle positioning system logic. A binocular calibration camera monitors the position of traffic lights in real time using ground targets. Deviations exceeding ±5 cm in height or 10 cm in lateral movement trigger the compensation mechanism, ensuring spatial consistency of light projection. A wide-angle monitoring camera detects the ellipticity of the light spot and automatically corrects the pitch angle to within 5% of the circularity error if the deviation exceeds the standard, eliminating interference from optical distortion on the perception system. The GNSS timing module achieves microsecond-level time synchronization using PPS pulses and the IEEE1588v2 protocol, reducing clock drift between the perception sensor and the control bus to less than 10 microseconds. The sliding time window algorithm, centered around the actual signal light switching moment t0, matches the response peak within a 700 millisecond window and automatically removes data with a time difference exceeding 80 milliseconds, ensuring the timing accuracy of causal analysis. The power isolation protocol disconnects the LED driver 5 milliseconds before injecting the simulated signal. The current sensor verifies zero current before injecting the signal, completely isolating the system from physical light interference. A 200 millisecond power delay mechanism avoids the vehicle response window and ensures the purity of the virtual test data.

[0027] Third, the speed weighting factor reduces the acceleration threshold to ±0.3 m / s² when exceeding 60 km / h, adapting to the braking characteristics of high-speed scenarios. A gradient release mechanism mitigates the impact of forced green light cuts by maintaining a 5-second interval followed by a 3-second yellow flash. Microwave detectors monitor the number of vehicles in the queue and the waiting time. When the threshold is exceeded, the light is cut to green and a priority pass code is issued, improving the efficiency of multi-vehicle coordination. Atomization equipment generates artificial fog with controllable concentration. Combined with a 50,000 to 100,000 lux adjustable light source, it reliably reproduces strong glare and low visibility scenarios, establishing a standardized test baseline for extreme environments. An ambient light sensor drives the brightness adaptation mechanism. Below 100 lux, the brightness is increased to 180%, addressing low-light recognition failures. Above 80,000 lux, a strobe mode is activated to suppress overexposure in strong light, expanding the coverage of test environments. The performance evaluation engine calculates latency mean and variance based on timestamp-aligned data streams. Tests exceeding 150 milliseconds latency or a squared variance of 30 milliseconds are automatically flagged as invalid. Real-time alerts are sent to the monitoring platform to ensure the reliability of the output dataset.

[0028] Other advantages, objectives and features of the present invention will be reflected in part through the following description, and in part will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Schematic diagram of the flow of the test method for the autonomous driving vehicle to recognize and respond to traffic lights in a closed area in the present invention. DETAILED DESCRIPTION

[0030] The present invention is described in further detail below so that those skilled in the art can implement the invention with reference to the description.

[0031] It should be understood that terms such as “having”, “including” and “comprising” used herein do not preclude the existence or addition of one or more other elements or combinations thereof.

[0032] It should be noted that the experimental methods described in the following embodiments are conventional methods unless otherwise specified, and the reagents and materials are commercially available unless otherwise specified. In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "set" should be understood in a broad sense, for example, they can be fixedly connected or set, or detachably connected or set, or integrally connected or set. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances. The orientations or positional relationships indicated by the terms "transverse", "longitudinal", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc. are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description. They do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention.

[0033] like Figure 1 As shown, the present invention provides a method for testing the recognition and response of an autonomous driving vehicle to a signal light in a closed area, comprising: S1. Digital construction of road network: Divide the roads in the closed site into line segments with directional attributes. Each line segment is defined by at least two geographic coordinate nodes. Adjacent line segments are connected to form a topological network by sharing nodes. S2. Dynamic traffic light deployment: deploying three-dimensional adjustable traffic light devices at the nodes of the topological network to adjust the suspension height, lateral offset and pitch angle in real time; S3, Multimodal Data Acquisition: Acquire signal light recognition data through vehicle-mounted perception sensors and simultaneously record vehicle control response instructions; S4, light color switching control: Based on the vehicle's positioning speed and the real-time distance from the stop line, the time required for the vehicle to reach the stop line is calculated as the time margin. The time margin is compared with a preset threshold. When the time margin meets the threshold condition, the signal light state is triggered to switch; S5. Two-dimensional verification: Synchronously collect the vehicle's traffic light recognition result data and control response behavior data. The recognition result data includes the traffic light status classification, spatial positioning coordinates and confidence level output by the perception system; the control response behavior data includes the brake pressure, throttle opening and steering angle instructions of the vehicle actuator. The two types of data are aligned through a time synchronization protocol to verify the causal relationship between the recognition result and the control response.

[0034] In the above embodiment, a typical intersection area was selected within a closed test field, and a road network was first digitally constructed. High-precision differential GPS was used to collect the geographic coordinates of each entrance lane centerline. Each lane centerline was defined as an independent line segment with a driving direction attribute, with the segment endpoints serving as geographic coordinate nodes. Each stop line at the intersection was designated as a key node, and adjacent lane segments were connected via shared nodes to form a complete topological network. A three-dimensional adjustable signal light device was then deployed above the corresponding stop line nodes. Exemplarily, this device includes a signal light body and a support structure. The support structure includes a lifting mechanism for adjusting the suspension height, a translation mechanism for adjusting the lateral offset, and a rotation mechanism for adjusting the pitch angle. The signal light body, serving as the terminal actuator, is mounted on the rotation mechanism, which itself is mounted on the translation mechanism, which in turn is mounted on the movable component of the lifting mechanism. For example, the lifting mechanism uses an electric push rod with a travel of 0-8m and an accuracy of ±1cm; the translation mechanism uses a servo motor-driven ball screw slide with a travel of ±2m and an accuracy of ±5mm; the rotation mechanism uses a worm gear reduction motor with a pitch adjustment range of -30° to +30° and an accuracy of ±0.5°. The supporting structures are all implemented using existing technologies, and as long as they can achieve the above functions, they will not be elaborated here. Therefore, the spatial position and orientation of the traffic light can be changed by real-time adjustment of the lifting, translation, and rotation mechanisms.

[0035] The test vehicle is equipped with a forward visual perception system and a controller area network bus. During the vehicle's driving process, the camera outputs the signal light status classification, spatial positioning coordinates and recognition confidence data in real time, and simultaneously collects the brake pressure command, throttle opening command and steering angle command issued by the vehicle controller. The roadside edge computing unit continuously receives vehicle positioning information, calculates the real-time distance between the vehicle and the stop line based on the preset stop line position, and calculates the time margin required to reach the stop line based on the vehicle's instantaneous speed. When the time margin drops to a fixed threshold T Y =2.5s, if the current light is green, the signal light immediately switches to a flashing yellow state. After three seconds of continuous flashing, the light automatically switches to red. When the time margin drops to (or below) the preset threshold T1, the light switches to red, regardless of the current light color. All sensing data and control commands are transmitted via optical fiber to a central server. The server's built-in precision clock source adds microsecond-level timestamps to both data streams, ensuring precise alignment between the signal light recognition result and the vehicle control response.

[0036] The closest existing technology uses a fixed testing framework: the traffic light is mechanically fixed at a height of six meters above the stop line, and the light color switching depends on a preset timing program, completely decoupled from the vehicle's motion state. The traffic light recognition results output by the vehicle's camera are stored in a local recorder, while the control response behavior data such as brake pedal travel is recorded in an independent controller. The two types of data rely only on the rough alignment of the vehicle's local clock, and the time synchronization error often exceeds 50 milliseconds. This method has significant limitations: statically installed traffic lights cannot simulate the perspective offset and optical distortion caused by vehicle trajectory changes in real intersections, resulting in a lack of verification of the perception algorithm's recognition ability under dynamic perspectives. The time misalignment between the traffic light recognition result data and the control response behavior data makes it impossible for testers to determine whether the root cause of the braking delay in the yellow light scenario is the perception system's recognition lag or the decision module's sluggish response, making it difficult to locate the failure link in the perception decision chain.

[0037] In one specific embodiment, in step S4, the signal light switching is triggered in a graded manner according to the time margin: S401, basic switching threshold: When the time margin is ≥ T2, it is forced to switch to the green light state; When the time margin is ≤ T1, it is forced to switch to the red light state; When T1 < time margin < T2, maintain the current light color state; where T1 and T2 meet the constraints: T2>3.0s>2.5s>T1; S402, Yellow Light Trigger Mechanism: The yellow light determination operation is performed when the following two conditions are met: i) the current time margin is less than or equal to 2.5 seconds; ii) the time margin changes from a value greater than 2.5 seconds, or it is the first time since system initialization that the time margin enters the range greater than T1 and not exceeding 2.5 seconds; The yellow light judgment operation is as follows: if the current light state is green, the green light state will be immediately switched to a flashing yellow light state, and the yellow light will automatically switch to a red light after flashing for 3 seconds; if the current light state is red or has already been yellow, the current state will remain unchanged; Each test vehicle only performs the yellow light determination operation once in a single test; while the yellow light is flashing, the execution of the S401 basic switching rule is suspended; when the vehicle passes the stop line, the yellow light determination state is reset; S403, Yellow light trigger sensitivity virtual test: During the yellow light flashing triggered by step S402, the actual signal light power is turned off and the vehicle communication interface is used to send The autonomous driving system injects a simulated signal; The analog signal states that the yellow light triggering time margin is T V , where T V∈ {2.6s, 2.7s, 2.8s, 2.9s,3.0s}; Traverse all T V The value is set and the vehicle response behavior is recorded. After the test, the power supply to the signal light is restored; S404, Dynamic Intervention Rules: When the basic switching interval (T1, T2) is within the range and the basic switching threshold is not violated, the following interventions are performed: i) If the vehicle acceleration is less than -0.5m / s 2 , when braking suddenly, the traffic light will be forced to turn red and the red light time will be extended; ii) If acceleration > 0.5 m / s 2 , then the current status is switched to green; iii) If any state lasts for more than 10 seconds, the light will be forced to turn green.

[0038] In the above embodiment, in the closed test field intersection test, the roadside control unit continuously calculates the time margin for the test vehicle to reach the stop line. When the time margin is greater than or equal to the preset threshold value T2, for example, T2 is set to 3.1 seconds, the traffic light is forced to switch to the green light state. When the time margin is between the preset threshold values ​​T1 and T2, for example, T1 is set to 1.8 seconds, the current state of the traffic light is maintained unchanged. When the time margin drops to less than or equal to 2.5 seconds, if the current traffic light is in the green light state, it will be immediately switched to a flashing yellow light state, and the yellow light will automatically turn to a red light after flashing for 3 seconds; if it is already a red light state at this time, the red light will remain unchanged. When the time margin is less than or equal to the preset threshold value T1, regardless of the current light color state, it will be forced to switch to a red light state.

[0039] During the 3-second period after the yellow light flashing state is triggered, the system performs a virtual test of the yellow light trigger sensitivity. At this time, the physical power supply of the actual signal light is automatically cut off, and at the same time, a simulated signal light status message is injected into the autonomous driving system through the vehicle communication interface. The simulated signal states that the time margin for triggering the yellow light is a specific test value T V , for example, T V The system then cycles through different time margins, such as 2.6, 2.7, 2.8, 2.9, and 3.0 seconds, and records the vehicle's response. After the entire virtual test sequence is complete, the signal light's power supply is restored.

[0040] In the basic switching interval, that is, when the time margin is between T1 and T2 and the basic switching threshold is not triggered, the system executes the dynamic intervention rule. 2 , it is judged as emergency braking behavior, at this time the red light is forced to switch and the red light duration is extended. If the vehicle acceleration exceeds 0.5 m / s 2, the current light color will be overwritten and switched to green directly. If any light color lasts for more than 10 seconds, the system will forcibly switch it to green to ensure traffic efficiency.

[0041] During the execution of dynamic intervention rules, if multiple intervention conditions are triggered at the same time (for example, the vehicle needs to switch to red light due to sudden braking and the traffic light needs to switch to green light due to timeout), the system will arbitrate according to the hierarchical principle of safety first and efficiency second: First, the vehicle is detected to have sudden braking behavior (i.e., the acceleration is less than -0.5 m / s 2 ), the light will be forced to switch to red immediately and the red light duration will be extended, while all other intervention rules will be interrupted; if the emergency braking condition is not triggered but the signal light state is detected to last for more than 10 seconds, the number of vehicles queued behind the stop line will be obtained through the microwave detector. When the number of vehicles in the queue is greater than or equal to 3, the light will be forced to switch to green; when the number of vehicles in the queue is less than 3, the current state will be maintained until the maximum tolerance time of 15 seconds before switching to green. After any rule is triggered, the system automatically enters a 5-second state lock cycle, during which new intervention instructions are blocked and status code C is output through the roadside display screen. lock =0xFF, 0xFF is the lock status identification code; after the lock cycle ends, the reset arbitrator resumes normal monitoring. For example, in the test, the vehicle acceleration reaches -0.6m / s 2 When a sudden braking action occurs simultaneously with a 12-second red light timeout, the system prioritizes the sudden braking to red light and activates a 5-second lockout, ignoring any timed-out switching requests. After the lockout expires, the system re-evaluates the switching request. This mechanism prevents signal state oscillation and keeps response latency within 100 milliseconds. By linking safety interrupts with traffic flow status, it ensures test safety and traffic flow consistency. The 5-second lockout period significantly reduces the system's computational load. When both an acceleration to green (rule ii) and a sudden braking to red (rule i) are triggered simultaneously, the sudden braking rule takes absolute priority, immediately terminating the acceleration to green.

[0042] The closest existing technology uses a fixed-time traffic light color switching scheme. Traffic light state changes strictly follow a pre-set schedule, for example, a green light lasting 30 seconds, then yellow for 3 seconds, and then red for 30 seconds. This scheme is completely detached from the real-time state of vehicles and cannot dynamically adjust the light color based on the actual time margin for vehicles approaching the intersection. In particular, when vehicles approach the stop line at varying speeds, the timing of encountering a yellow light is random and uncontrollable, making it impossible to accurately replicate the standardized yellow light decision scenario with a 2.5-second time margin.

[0043] Existing technologies lack the ability to specifically test sensitivity to yellow light decisions. During testing, the physical signal lights remain powered on, making it impossible to isolate the interference of actual lights on the perception system. Furthermore, there is no capability to inject simulated signals with a variable time margin of 2.6 to 3.0 seconds into the vehicle. Its dynamic intervention mechanism relies solely on a simple timer, unable to respond in real time to sudden braking for red lights or accelerating for green lights based on the vehicle's acceleration direction. Nor does it have traffic optimization logic to force a green light switch if a persistent state timeout occurs.

[0044] In one specific embodiment, a positioning error compensation mechanism is added to the time margin calculation in step S4: The vehicle positioning correction vector is generated in real time by a high-precision RTK differential base station deployed on the roadside. The correction vector includes the horizontal position offset (δ X , δ Y ) and heading angle deviation δθ; the original coordinates (X V , Y V ) According to the formula (X V +δ X ·cosδθ,Y V +δ Y sinδθ) for dynamic calibration; At the same time, an elastic buffer model for the stop line position is established. In the stop line elastic buffer model, the boundary range is generated by the following formula: gray decision interval half-width W = max(0.6m, 3σ), where σ is the standard deviation of the horizontal precision factor of the current RTK differential base station positioning. When the actual distance D between the calibrated vehicle coordinates and the stop line real , satisfying |D real -D calc When |>0.5m, the virtual boundary of the stop line is automatically extended to ±W, D calc The original calculated distance from the stop line reported by the vehicle positioning system; If the vehicle position does not cross the virtual boundary, that is, in the interval [D calc -W,D calc +W], the signal light status switches according to the S4 rule, and the dual-channel data recording mechanism is started, specifically: i) First channel recording: the original calculated distance D calc The vehicle's recognition and response behavior to the signal light based on the actual distance D real The vehicle's recognition and response behavior to traffic lights based on the baseline; When the vehicle position crosses the virtual boundary, the second channel data is automatically discarded and the first channel data is output as the valid test result.

[0045] In the above implementation, a high-precision RTK differential base station can be installed on the roadside. The equipment can use a GNSS receiver that supports multi-band satellite signals. The housing material can be made of aluminum alloy to reduce weight. The base station is usually fixed on the top of the light pole in the test site or on the roof of the building, with an installation height of not less than 6 meters. The base station monitors the original coordinates of the vehicle (X V， Y V ), generates a matrix containing the horizontal offset δ X , δ Y And the correction vector of the heading angle deviation δθ. X , δ X The value range is usually within ±1.0 meters, and the range of δθ is within ±5 degrees. The correction vector is transmitted to the vehicle system via the 5G wireless module at a frequency of 10 times per second.

[0046] After receiving the correction vector, the vehicle positioning system processes it according to the dynamic calibration rules: the original coordinates (X V， Y V ) is converted to calibration coordinates (X V + δ X ·cosδθ,Y V + δ Y ·sinδθ). The elastic buffer model is set at the stop line position, the boundary half width W is taken as max(0.6m, 3σ), and the standard deviation σ value can be set to 0.15m to 0.25m. When the actual distance D real Distance D from the original calculation calc When the deviation exceeds 0.5 meters, the virtual boundary of the stop line is automatically expanded to the range of ±W. The virtual boundary data is stored in the memory of the roadside edge computing device.

[0047] When the actual distance D between the calibrated vehicle coordinates and the stop line real , satisfying |D real -D calc When |>0.5m, the virtual boundary of the stop line is automatically extended to ±W. At this time, the signal light state switching strictly follows the original calculated distance D of the first channel. calc Control logic, and D real If the vehicle position does not cross the virtual boundary (ie D real ∈ [ D calc -W , D calc + W], the signal light status switches according to S4 rule, and dual-channel data recording is started: the first channel: D calc As the benchmark; the second channel: D real As the reference. When the vehicle position crosses the virtual boundary (ie D real Does not belong to [D calc -W , D calc+ W] ) :i) Immediately stop the second channel data recording; ii) The signal light control is completely transferred to the first channel logic, and continues to calc is the reference switching state; iii) only the first channel data is used as the valid output.

[0048] This solution reduces false triggering of signal lights through real-time positioning calibration and utilizes a flexible buffer model to adapt to positioning scenarios with varying degrees of accuracy. Its dual-channel data recording mechanism retains valid test data even in complex electromagnetic environments, improving the reliability of closed-site testing.

[0049] For example, in a closed field, the test section below the viaduct is blocked by satellite signals, and the original coordinates reported by the vehicle positioning system are 15.6 meters away from the stop line. At this time, the roadside RTK differential base station detects a horizontal position offset δ X 0.3 m, δ Y The actual distance D is generated by the dynamic calibration formula. real is 16.2 meters; since this value is different from the original calculated distance D calc If the deviation reaches 0.6 meters and exceeds the 0.5-meter threshold, the system automatically expands the virtual boundary of the stop line and calculates the elastic buffer half-width W=0.6 meters based on the current RTK positioning precision factor standard deviation σ=0.2 meters; at this time, although the vehicle has not crossed the virtual boundary but is within the boundary range of 15.0 meters to 16.2 meters, the system simultaneously starts dual-channel data recording: the first channel records the vehicle recognition response behavior based on the original calculated distance of 15.6 meters, and the second channel records the response behavior based on the actual distance of 16.2 meters; when the vehicle continues to move forward and the actual position exceeds the 16.2-meter virtual boundary, the system immediately discards the second channel data and only retains the valid test results of the first channel.

[0050] Compared to the closest existing technology, the publicly available test site positioning compensation solution only uses a single GPS offset correction. This caused positioning drift when the vehicle entered a tunnel, resulting in a calculated distance of 12 meters from the stop line, while the actual distance was only 8 meters. The system's lack of an elastic buffer model directly triggered a red light switch, forcing the autonomous driving system to perform emergency braking and interrupt the test. This solution uses RTK to generate a 3D correction vector in real time, combined with a stop line elastic buffer mechanism and dual-channel data recording, to avoid false triggering of signal switches while preserving valid test data, significantly improving test continuity in complex scenarios.

[0051] In one specific embodiment, after the dynamic traffic light is deployed in step S2, a spatial posture closed-loop verification needs to be performed: A reference target pattern is laid on the ground directly below the traffic light, and the relative position of the traffic light and the target is captured by a binocular calibration camera mounted on top of the test vehicle. When it is detected that the actual hanging height H of the signal light deviates from the set value by more than ±5cm, or the lateral offset L is greater than 10cm, the posture compensation mechanism is automatically triggered to perform posture compensation; At the same time, the optical distortion detection module is activated after each light color change. Specifically, a wide-angle monitoring camera installed on the back of the traffic light is used to capture the shape of the light spot projected by the light. If the light spot ellipticity ε>0.25, the pitch angle is automatically corrected until the roundness error ρ≤5%.

[0052] In the above embodiment, a reference target is laid on the ground directly below the traffic light. The target can be made of a checkerboard pattern of high-contrast reflective material with a size of 1 meter × 1 meter. A binocular calibration camera can be installed on the top of the test vehicle, and the camera bracket is 2.5 meters from the ground. When the traffic light is on, the binocular camera captures the relative position image of the traffic light and the target at a rate of 30 frames per second, and calculates the actual hanging height H and lateral offset L of the traffic light through a visual algorithm. If it is detected that the deviation between H and the set value exceeds ±5 cm or L is greater than 10 cm, the posture compensation mechanism is triggered. The compensation mechanism can use an electric push rod, which is installed at the adjustment node of the traffic light support rod.

[0053] The posture compensation mechanism consists of three sets of electric actuators, controlling height, lateral displacement, and pitch angle, respectively. The actuators can be made of stainless steel, with a travel range of ±20 cm in height, ±15 cm in lateral displacement, and ±10 degrees in pitch angle. After the compensation action is executed, the system recaptures the target image for secondary verification. After each light color change, the light spot is captured by a wide-angle monitoring camera on the back of the signal light. The camera, an industrial model with a 120-degree field of view, is mounted in the center of the back panel of the light housing. If the light spot ellipticity ε exceeds 0.25, automatic pitch angle correction is initiated until the roundness error ρ does not exceed 5%. This correction process is controlled by a microprocessor, with a single cycle taking less than 200 milliseconds. The posture compensation mechanism is directly mounted on the rotating mechanism of the support structure. The electric actuators of the posture compensation mechanism are mechanically coupled to the worm gear reduction motor of the rotating mechanism. The posture compensation mechanism, as an integrated expansion module of the rotating mechanism, is rigidly connected via flange bolts.

[0054] The workflow is as follows: After signal light initialization, a binocular camera performs an initial position measurement. Any deviation exceeding the specified tolerance triggers a compensation mechanism to adjust the position. When the light changes color, a wide-angle camera simultaneously initiates spot analysis. Any ellipticity deviation detected is immediately fed back to the pitch angle controller. All image processing utilizes an embedded GPU module installed in the roadside control box. Verification data is stored in real time, including the displacement of each compensation, the corrected ellipticity value, and a timestamp. Absolute spatial position (height / offset) is first verified using binocular calibration with a ground target. Any deviation triggers mechanical compensation. After each light color change, a rear-facing wide-angle camera detects spot distortion. Pitch angle correction is only performed when ellipticity ε exceeds 0.25, creating a dual-layered guarantee: "basic positioning → auxiliary lighting effect."

[0055] This implementation achieves millimeter-level pose verification through a combination of a target and dual cameras, while an automatic compensation mechanism maintains the stability of signal light spatial parameters. Real-time optical distortion detection ensures the correct projection of light, preventing interference caused by installation offset or vibration, and improving the standardization of test scenarios.

[0056] In one specific embodiment, the dual-dimensional verification in step S5 requires the establishment of a hardware-level time synchronization system, specifically: A GNSS timing module is installed in the traffic light control cabinet to generate PPS pulse signals, which are distributed to onboard sensing sensors and the vehicle control bus via a fiber optic splitter. All data streams are tagged with the IEEE 1588v2 protocol, including a UTC microsecond timestamp. To align the timing of signal light recognition data with control response instructions, a sliding time window matching algorithm is set: based on the actual signal light switching time t0, the vehicle response data peak is searched in the interval [t0-200ms, t0+500ms]. When the time difference Δt between the recognition result and the response behavior is greater than 80ms, the test data is automatically discarded and the test sequence is retriggered.

[0057] In the above embodiment, a GNSS timing module can be installed in the traffic light control cabinet. The module can use a multi-constellation receiving chip, and the shell material can be made of aluminum alloy. The antenna of the GNSS module can be installed in an open-air location on the top of the control cabinet. The module generates a pulse-per-second PPS signal with an accuracy of up to 1 microsecond. The PPS signal is distributed through a fiber optic splitter. The splitter can use a 1-to-8-way single-mode optical fiber type and be installed in the wiring rack of the traffic light control cabinet. The optical fiber line can use a 9 / 125μm single-mode jumper, which is respectively connected to the timing interfaces of the on-board camera, lidar and other perception sensors and the vehicle control bus. All timing interfaces can use SFP optoelectronic conversion modules, which are installed on the communication backplane of the on-board equipment box.

[0058] All signal light recognition data (such as light color classification and location coordinates) collected by the perception sensors, as well as control response data (such as brake pressure commands and steering angle commands) collected by the vehicle bus, are tagged with the IEEE 1588v2 protocol. The tag generation device can use a switching chip that supports the precision clock protocol and be integrated into the vehicle's industrial computer motherboard. The tag is embedded with a UTC timestamp, which is locked to the PPS signal output by the GNSS module, with a timestamp accuracy of 1 microsecond. When data streams are transmitted over the in-vehicle Ethernet, clock de-skew is performed every 200 milliseconds, with a de-skew threshold set to 50 nanoseconds. If the sensor and bus clocks deviate by more than 10 microseconds, a hardware reset signal is automatically triggered.

[0059] Using the actual signal light switching time t0 as the reference point, the central processing unit initiates a sliding time window matching algorithm. The time window range is fixed from 200 milliseconds before t0 to 500 milliseconds after t0, for a total of 700 milliseconds. Within this window, a search is performed for vehicle response data peaks, such as the maximum slope of the brake pressure curve or the sudden change in throttle opening. Peak detection uses a gradient ascent algorithm, with a peak determined when the gradient values ​​of three consecutive points decrease. If the difference Δt between the timestamp of the recognition result data and the timestamp of the response behavior data is greater than 80 milliseconds, the system automatically discards all data from that test sequence. After discarding, the test process is immediately retriggered, resetting the signal light to its initial state. The data screening results are recorded on a solid-state storage disk, which can use an industrial-grade SATA interface model and be installed in the hard drive slot of the roadside cabinet.

[0060] This solution eliminates clock drift in multi-source data through a hardware-level time synchronization system, ensuring precise timing correlation between perception and response data. A sliding time window algorithm effectively captures key vehicle behavioral characteristics after signal switching. A data discard mechanism filters out test samples with misaligned timing, improving the reliability of causal analysis results.

[0061] In one specific embodiment, a power isolation protocol is executed during the yellow light trigger sensitivity virtual test in step S403: i) 5 milliseconds before injecting the simulated signal, the signal light LED driver power is cut off; ii) the simulated signal is injected again after verifying that the current drops to 0A through a current sensor; iii) at the end of the test, the power supply is restored with a delay of 200 milliseconds to avoid the vehicle response window.

[0062] In the above embodiment, 5 milliseconds before the analog signal is injected, the system cuts off the power to the signal light LED driver. The power cutoff module can be a solid-state relay, packaged in a ceramic substrate with a rated current of 60A. This relay is mounted on the driver board of the signal light control cabinet and receives the trigger signal via an optocoupler isolation circuit. After the cutoff command is issued, a current sensor monitors the LED circuit current in real time. The sensor can be a closed-loop Hall effect type with a range of 0-10A and an accuracy of ±10mA, and is mounted at the output terminal of the driver board. When the current drops to 0A and remains at 0A for 1 millisecond, a current zeroing signal is fed back to the main control unit. After the main control unit confirms that the current has returned to zero, it injects an analog signal through the vehicle's communication interface. The communication interface can use a CAN FD bus transceiver, and the material can be an FR-4 epoxy resin substrate, which is installed in the vehicle's OBD-II port expansion slot. The analog signal includes the yellow light trigger time margin T V Parameter, T V The value ranges from 2.6 seconds to 3.0 seconds in 0.1 second increments. The injection process lasts 20 milliseconds, and the data packet is repeated three times to ensure reception. The signal generation device can use an embedded microcontroller and is installed in a slot in the roadside cabinet. The operating temperature range is -40°C to 85°C. At the end of the test, the system delays 200 milliseconds to restore the power supply to the signal light. The delay control can use a programmable timer chip with a timing error of less than 1 millisecond. The material can be a flame-retardant plastic shell, which is installed in the power management unit of the control cabinet. Before restoring power, the vehicle status monitoring module confirms that the response window has ended. The monitoring module can use an on-board accelerometer with a range of ±10g, which is installed in the center of the vehicle chassis. If the acceleration change rate is continuously lower than 0.05g / s within 50 milliseconds 2 , it is determined that the vehicle response has ended and the power supply recovery instruction is triggered.

[0063] This protocol completely isolates physical lighting interference through precise power cutoff and current verification, ensuring the purity of virtual test signals. A 200-millisecond power delay effectively avoids vehicle response windows and prevents test results from being contaminated by actual lighting colors. Real-time current monitoring and timer control enhance system reliability.

[0064] In one specific embodiment, in step S404 of step S4, a vehicle speed weighting factor is added to the dynamic intervention rule: When the vehicle speed v>60km / h, the acceleration threshold a th Adaptive adjustment is ±0.3m / s²; For the rule of forcing the light to turn green in a continuous state, a gradient release mechanism is set: the current light color is maintained for the first 5 seconds, a yellow warning flash is started at the 6th to 8th second, and the light switches to green at the 9th second; At the same time, microwave detectors are deployed at the topological network nodes to monitor the queue length. If the number of vehicles queuing behind the stop line N ≥ 3 and the average waiting time T av >8 seconds, automatically override the acceleration judgment rule and switch to green light directly, and push the priority passage prompt code C on the roadside display priority .

[0065] In the above embodiment, when the vehicle speed exceeds 60 km / h, the system adaptively adjusts the acceleration threshold to ±0.3 m / s². Speed ​​detection can be performed using a Doppler radar sensor, constructed from a polycarbonate housing and mounted in the center of the vehicle's front bumper. Acceleration threshold adjustments are calculated in real time by an onboard microprocessor, which can utilize a 32-bit ARM Cortex-M7 core operating at 480 MHz and slotted into the vehicle's control unit. The adjusted threshold remains in effect until the vehicle speed drops below 55 km / h, with a hysteresis interval of 5 km / h to prevent frequent switching.

[0066] To ensure a continuous green light, the system implements a three-stage gradient release process: the current light color remains unchanged for the first five seconds, a yellow warning flashing (2Hz frequency) is activated from the sixth to eighth seconds, and the light switches to green at the ninth second. The countdown controller can utilize a real-time operating system timer module made of a ceramic oscillator and installed on the timing board of the roadside signal control cabinet. The yellow flash drive circuit utilizes a constant-current LED driver with an output current of 1.5A ± 5%, installed in the signal power supply box. Switching commands are transmitted via the RS-485 bus, with a baud rate set to 115200bps.

[0067] Deploy microwave detectors at topological network nodes. The detectors can be 24GHz frequency modulated continuous wave models with a detection range of 80 meters. They are installed 5.5 meters above the ground on the signal light bar. Real-time monitoring of the number of vehicles N and the average waiting time T in the queue behind the stop line av When N≥3 vehicles and T av When the speed is >8 seconds, the system overrides the acceleration judgment rule and switches the light to green directly. The roadside display screen will simultaneously push the priority passage prompt code C priority The display screen can use a high-brightness LED matrix module with a pixel pitch of 4mm and be installed on the roadside column 20 meters before the stop line. In the dynamic intervention rules, the priority pass prompt code C priority This code uses a 1-byte hexadecimal encoding (e.g., 0x5A), with bits divided into three layers of information: the high-order bits (bits 7-6) identify the intersection topology node ID (e.g., 01 for node A); the middle bits (bits 5-4) define the priority type (bit 10 is specifically for test vehicles to have priority); and the low-order bits (bits 3-0) carry dynamic parameters (e.g., 1010 for a 10-second countdown). When the roadside system detects a queue of ≥3 vehicles and an average wait time >8 seconds, it automatically generates this code and displays it on the LED display as a green pulse. If the autonomous vehicle interprets the code and matches its own test identity (bits 5-4 = 10), it triggers the collaborative algorithm to smoothly accelerate through the intersection at 90% of the speed limit, optimizing the passage sequence using the countdown parameters. This design ensures precise, machine-readable instruction transmission while also reserving expansion bits (e.g., bits 5-4 = 11) to accommodate future priority level expansion.

[0068] Optimized vehicle speed weighting factors adapt to high-speed braking characteristics, reducing the risk of misintervention. A gradient release mechanism mitigates the impact of light color changes by flashing yellow during transitions. Traffic flow coordination improves multi-vehicle testing efficiency, and priority pass codes provide visual indicators for subsequent algorithm analysis. Microwave detection and display devices collaborate to enhance system environmental adaptability.

[0069] In one specific embodiment, an environmental interference simulation module is added to the multimodal data acquisition in step S3: Through atomization equipment, artificial fog of controllable concentration is generated in front of the traffic light, and the ambient light intensity is simultaneously adjusted to 50,000-100,000 lux to simulate strong glare scenes, and the vehicle's signal light recognition data under extreme conditions is collected.

[0070] In the above embodiment, a controllable concentration of fog is generated in front of the traffic light by an atomizing device, and the fog concentration range is set to 0.1 to 0.3 grams per cubic meter. The atomizing device can use an ultrasonic atomizer array, with a single-machine atomization volume of 300 milliliters per hour. The material can be a 316 stainless steel atomizer sheet, which is installed on a roadside column 3 to 5 meters away from the traffic light. The column height is 1.2 meters, and the atomizing nozzle is tilted upward at an angle of 30 degrees. Concentration control is achieved by adjusting the PWM duty cycle, with a duty cycle step of 5% and a response time of less than 200 milliseconds. The environmental humidity sensor monitors the fog diffusion status in real time. The sensor can use a capacitive polymer film model and is installed 2 meters directly in front of the traffic light and 2.5 meters above the ground.

[0071] Ambient light intensity is synchronously adjusted to a range of 50,000 to 100,000 lux. The light source can be a xenon lamp assembly, with a single lamp power of 1,000 watts and a color temperature of 5,500K ± 300K, made of quartz glass tubes. The lamp assembly is mounted 10 meters behind the signal light on a retractable bracket, with an adjustable height range of 2 to 6 meters. Light intensity is adjusted using a digital dimmer, which can be a silicon-controlled rectifier module with a current accuracy of ±0.5 amperes and is installed in the roadside distribution box. Illuminance calibration is performed using a secondary standard light meter, with the measuring probe fixed to the center of the vehicle camera's field of view, 1.5 meters above the ground.

[0072] After starting the test sequence, the system first activates the atomization device for 30 seconds to form a stable mist layer. Simultaneously, the illumination of the xenon lamp cluster is linearly increased to the target value (e.g., 80,000 lux) over 10 seconds. The onboard perception sensor collects signal light recognition data at 25 frames per second, focusing on the following parameters: color classification confidence, bounding box positioning accuracy, and recognition latency. Data is transmitted via an interference-resistant shielded cable. The cable can be a twisted-pair, tinned copper core with PVC insulation, with the wiring path fixed at 0.5 meters above the ground. Each test lasts 120 seconds. After completion, all interference sources are turned off and the environment is ventilated for 5 minutes to reset.

[0073] This module precisely controls fog concentration and light intensity parameters to stably reproduce low-visibility and high-glare scenarios. Standardized testing procedures ensure comparability of data across batches, providing a quantifiable environmental benchmark for validating the robustness of perception algorithms under extreme conditions. Equipment deployment locations and parameter thresholds are designed to comply with safety regulations for traffic test sites.

[0074] In one specific implementation, a brightness adaptation mechanism is added after the spatial pose closed-loop verification, specifically: An ambient light sensor is integrated into the signal light control cabinet to monitor the ambient illumination in real time. When the ambient illumination is lower than 100 lux, the LED drive current is automatically increased to increase the light brightness to 180% of the standard value. When the ambient illumination is higher than 90,000 lux, the pulse strobe mode of the signal light is activated.

[0075] In the above embodiment, an ambient light sensor can be integrated into the traffic light control cabinet. The sensor can be a silicon photocell with a spectral response range of 380 to 720 nanometers, and can be encapsulated in epoxy resin. The sensor is mounted on the front ventilation hole of the control cabinet, 1.6 meters above the ground, with the photosensitive surface tilted upward at 15 degrees to avoid direct sunlight. It monitors ambient illumination in real time, sampling at a frequency of 10 times per second. If the sampled values ​​are below 100 lux for five consecutive times, the low-light response mode is triggered; if the sampled values ​​are above 90,000 lux for five consecutive times, the high-light response mode is triggered.

[0076] In low-light mode, the system automatically increases the LED drive current to 180% of the standard value. The drive circuit can use a constant current source chip with an adjustable output current range of 1 to 3 amps. The material can be an alumina ceramic substrate, which is installed on the heat sink surface of the signal light power module. The current increase process uses ramp control, linearly increasing to the target value within 200 milliseconds to avoid current surges. Brightness calibration is verified through a closed-loop photometric probe. The probe can use a secondary standard light meter, which is temporarily fixed 10 meters in front of the signal light and 1.2 meters above the ground. After calibration, it is removed.

[0077] In strong light mode, the pulse strobe function is activated. The strobe controller can use a programmable logic device with an output frequency of 100 Hz square wave and a duty cycle of 30%±5%, and is installed in the expansion slot of the driver board. The LED lamp beads work under pulse drive, and the peak brightness is increased to 250% of the standard value, using the human eye's visual persistence effect to maintain equivalent brightness. The optical lens can use a polycarbonate Fresnel pattern model with a thickness of 3 mm, which is installed on the inside of the lampshade to disperse strong light and avoid overexposure. The mode switching delay is controlled within 50 milliseconds, and the junction temperature of the lamp beads is monitored by a temperature sensor. When it exceeds 85°C, the frequency is automatically reduced to 80 Hz. This mechanism ensures signal visibility in low-light conditions by monitoring ambient illumination in real time and responding dynamically to it. It also prevents image overexposure in strong glare. Current ramp control and temperature monitoring ensure safe operation, while a strobe mode combined with optical design enhances recognition stability in extreme lighting conditions. Mounting location and threshold settings comply with road lighting standards.

[0078] In one embodiment, a performance evaluation engine is deployed after the two-dimensional verification. Specifically, based on the data stream after time stamp alignment, the mean and variance of the delay from signal light recognition to vehicle control response are calculated. When the delay exceeds 150ms or the variance is greater than 30ms, the vehicle control response is automatically generated. 2 When the test fails, the test is automatically marked as invalid and an alarm signal is triggered to the monitoring platform.

[0079] In the above embodiment, based on the data stream after timestamp alignment, the system calculates the delay mean and variance from signal light recognition to vehicle control response. The computing device can use an industrial edge computing unit, the processor can use a quad-core ARMCortex-A72 architecture with a main frequency of 1.8GHz, and the material can use an aluminum alloy heat dissipation shell, which is installed on the upper guide rail of the roadside cabinet. The data input interface is a dual-channel Gigabit Ethernet, and the cache capacity is set to 8GB. The delay calculation is based on a single test as a unit, and the sampling window contains 50 consecutive groups of data points. The mean adopts the arithmetic mean algorithm, and the variance calculation uses the standard unbiased estimation formula. The calculation cycle is fixed to be executed once every 200ms. Set the delay mean threshold to 150ms and the variance threshold to 30ms 2 The judgment module can use a programmable logic chip with a comparator accuracy of 1 nanosecond and is installed in the expansion slot of the edge computing unit. When the delay mean exceeds 150ms or the variance is greater than 30ms, 2 When an invalid test occurs, an invalid test marker is automatically generated. The marker data packet contains the test ID, the expiration timestamp, and the exceeded parameter value. The data packet structure is a 32-byte fixed-length format. The marker record is written to the database in real time. The database can use time-series database software and is stored on a 2TB solid-state drive mounted on the cabinet's shock-absorbing bracket.

[0080] Once an invalid flag is generated, an alarm signal is immediately triggered to the monitoring platform. Three alarm channels are available: the SMS alarm module can use a 4G Cat1 communication module and be installed in the cabinet's communication compartment; the audible and visual alarm can use a 120dB buzzer and red LED combination and be installed inside the cabinet's front door; and the platform interface uses the RESTful API protocol and runs on the monitoring center server. The alarm content includes the test location code, failure type (mean or variance exceeded), and recommended action. The alarm persists until manually acknowledged. Acknowledgment is issued via a button on the monitoring platform, with a response latency requirement of less than 500ms.

[0081] The engine automatically identifies abnormal test data through real-time latency analysis, and an invalid marking mechanism prevents contamination of the resulting dataset. Multi-channel alarm linkage ensures rapid response from operations and maintenance personnel, improving the operational efficiency of the closed test site. Hardware deployment and threshold settings meet the reliability standards of intelligent transportation systems.

[0082] The number of devices and processing scales described herein are intended to simplify the description of the present invention. Applications, modifications, and variations of the present invention will be readily apparent to those skilled in the art.

[0083] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the description and implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to specific details.

Claims

1. A method for testing the recognition and response of autonomous vehicles to traffic lights in a closed area, characterized by: include: S1. Digital construction of road network: Divide the roads in the closed site into line segments with directional attributes. Each line segment is defined by at least two geographic coordinate nodes. Adjacent line segments are connected to form a topological network by sharing nodes. S2. Dynamic traffic light deployment: deploying three-dimensional adjustable traffic light devices at the nodes of the topological network to adjust the suspension height, lateral offset and pitch angle in real time; S3, Multimodal Data Acquisition: Acquire signal light recognition data through vehicle-mounted perception sensors and simultaneously record vehicle control response instructions; S4, light color switching control: Based on the vehicle's positioning speed and the real-time distance from the stop line, the time required for the vehicle to reach the stop line is calculated as the time margin. The time margin is compared with a preset threshold. When the time margin meets the threshold condition, the signal light state is triggered to switch; S5. Two-dimensional verification: Synchronously collect the vehicle's traffic light recognition result data and control response behavior data. The recognition result data includes the traffic light status classification, spatial positioning coordinates and confidence level output by the perception system; the control response behavior data includes the brake pressure, throttle opening and steering angle instructions of the vehicle actuator. The two types of data are aligned through a time synchronization protocol to verify the causal relationship between the recognition result and the control response.

2. The method for testing the recognition and response of an autonomous driving vehicle to a signal light in a closed area according to claim 1, wherein: In step S4, the signal light switching is triggered in a hierarchical manner according to the time margin, specifically: S401, basic switching threshold: When the time margin is ≥ T2, it is forced to switch to the green light state; When the time margin is ≤ T1, it is forced to switch to the red light state; When T1 < time margin < T2, maintain the current light color state; where T1 and T2 meet the constraints: T2>3.0s>2.5s>T1; S402, Yellow Light Trigger Mechanism: The yellow light determination operation is performed when the following two conditions are met: i) the current time margin is less than or equal to 2.5 seconds; ii) the time margin changes from a value greater than 2.5 seconds, or it is the first time since system initialization that the time margin enters the range greater than T1 and not exceeding 2.5 seconds; The yellow light judgment operation is as follows: if the current light state is green, the green light state will be immediately switched to a flashing yellow light state, and the yellow light will automatically switch to a red light after flashing for 3 seconds; if the current light state is red or has already been yellow, the current state will remain unchanged; Each test vehicle only performs the yellow light determination operation once in a single test; while the yellow light is flashing, the execution of the S401 basic switching rule is suspended; when the vehicle passes the stop line, the yellow light determination state is reset; S403, Yellow light trigger sensitivity virtual test: During the yellow light flashing triggered by step S402, the actual signal light power is turned off and the vehicle communication interface is used to send The autonomous driving system injects a simulated signal; The analog signal states that the yellow light triggering time margin is T V , where T V ∈ {2.6s, 2.7s, 2.8s, 2.9s,3.0s}; Traverse all T V The value is set and the vehicle response behavior is recorded. After the test, the power supply to the signal light is restored; S404, Dynamic Intervention Rules: When the basic switching interval (T1, T2) is within the range and the basic switching threshold is not violated, the following interventions are performed: i) If the vehicle acceleration is less than -0.5m / s 2 , when braking suddenly, the traffic light will be forced to turn red and the red light time will be extended; ii) If acceleration > 0.5 m / s 2 , then the current status is switched to green; iii) If any state lasts for more than 10 seconds, the light will be forced to turn green.

3. The method for testing the recognition and response of an autonomous driving vehicle to a signal light in a closed area according to claim 2, wherein: A positioning error compensation mechanism is added to the time margin calculation in step S4, specifically: The vehicle positioning correction vector is generated in real time by a high-precision RTK differential base station deployed on the roadside. The correction vector includes the horizontal position offset (δ X , δ Y ) and heading angle deviation δθ; the original coordinates (X V , Y V ) According to the formula (X V +δ X ·cosδθ,Y V +δ Y sinδθ) for dynamic calibration; At the same time, an elastic buffer model for the stop line position is established. In the stop line elastic buffer model, the boundary range is generated by the following formula: gray decision interval half-width W = max(0.6m, 3σ), where σ is the standard deviation of the horizontal precision factor of the current RTK differential base station positioning. When the actual distance D between the calibrated vehicle coordinates and the stop line real , satisfying |D real -D calc | > 0.5m, automatically expand the virtual boundary of the stop line to ±W, D calc The original calculated distance from the stop line reported by the vehicle positioning system; If the vehicle position does not cross the virtual boundary, that is, in the interval [D calc -W,D calc +W], the signal light status switches according to the S4 rule, and the dual-channel data recording mechanism is started, specifically: i) First channel recording: the original calculated distance D calc The vehicle's recognition and response behavior to traffic lights based on the baseline; ii) Second channel recording: actual distance D real The vehicle's recognition and response behavior to traffic lights based on the baseline; When the vehicle position crosses the virtual boundary, the second channel data is automatically discarded and the first channel data is output as the valid test result.

4. The method for testing the recognition and response of an autonomous driving vehicle to a signal light in a closed area according to claim 1, wherein: After the dynamic traffic light is deployed in step S2, a closed-loop verification of the spatial posture needs to be performed, specifically: A reference target pattern is laid on the ground directly below the traffic light, and the relative position of the traffic light and the target is captured by a binocular calibration camera mounted on top of the test vehicle. When it is detected that the actual hanging height H of the signal light deviates from the set value by more than ±5cm, or the lateral offset L is greater than 10cm, the posture compensation mechanism is automatically triggered to perform posture compensation; At the same time, the optical distortion detection module is activated after each light color change. Specifically, a wide-angle monitoring camera installed on the back of the traffic light is used to capture the shape of the light spot projected by the light. If the light spot ellipticity ε>0.25, the pitch angle is automatically corrected until the roundness error ρ≤5%.

5. The method for testing the recognition and response of an autonomous driving vehicle to a signal light in a closed area according to claim 1, wherein: The dual-dimensional verification in step S5 requires the establishment of a hardware-level time synchronization system, specifically: A GNSS timing module is installed in the traffic light control cabinet to generate PPS pulse signals, which are distributed to onboard sensing sensors and the vehicle control bus via a fiber optic splitter. All data streams are tagged with the IEEE 1588v2 protocol, including a UTC microsecond timestamp. To align the timing of signal light recognition data with control response instructions, a sliding time window matching algorithm is set: based on the actual signal light switching time t0, the vehicle response data peak is searched in the interval [t0-200ms, t0+500ms]. When the time difference Δt between the recognition result and the response behavior is greater than 80ms, the test data is automatically discarded and the test sequence is retriggered.

6. The method for testing the recognition and response of an autonomous driving vehicle to a signal light in a closed area according to claim 2, wherein: During the yellow light trigger sensitivity virtual test in step S403 , the power isolation protocol is executed: i) 5 milliseconds before injecting the simulated signal, the signal light LED driver power is cut off; ii) After verifying that the current has dropped to 0A using a current sensor, the analog signal is injected again. iii) At the end of the test, power is restored with a delay of 200 milliseconds to avoid the vehicle's response window.

7. The method for testing the recognition and response of an autonomous driving vehicle to a signal light in a closed area according to claim 2, wherein: In step S404 of step S4, a vehicle speed weight factor is added to the dynamic intervention rule, specifically: When the vehicle speed v>60km / h, the acceleration threshold a th Adaptive adjustment is ±0.3m / s²; For the rule of forcing the light to turn green in a continuous state, a gradient release mechanism is set: the current light color is maintained for the first 5 seconds, a yellow warning flash is started at the 6th to 8th second, and the light switches to green at the 9th second; At the same time, microwave detectors are deployed at the topological network nodes to monitor the queue length. If the number of vehicles queuing behind the stop line N ≥ 3 and the average waiting time T av >8 seconds, automatically override the acceleration judgment rule and switch to green light directly, and push the priority passage prompt code C on the roadside display priority .

8. The method for testing the recognition and response of an autonomous driving vehicle to a signal light in a closed area according to claim 1, wherein: In the multimodal data acquisition of step S3, an environmental interference simulation module is added, specifically: Through atomization equipment, artificial fog of controllable concentration is generated in front of the traffic light, and the ambient light intensity is simultaneously adjusted to 50,000-100,000 lux to simulate strong glare scenes, and the vehicle's signal light recognition data under extreme conditions is collected.

9. The method for testing the recognition and response of an autonomous driving vehicle to a signal light in a closed area according to claim 4, wherein: After the closed-loop verification of the spatial pose, a brightness adaptation mechanism is added, specifically: An ambient light sensor is integrated into the signal light control cabinet to monitor the ambient illumination in real time. When the ambient illumination is lower than 100 lux, the LED drive current is automatically increased to increase the light brightness to 180% of the standard value. When the ambient illumination is higher than 90,000 lux, the pulse strobe mode of the signal light is activated.

10. The method for testing the recognition and response of an autonomous driving vehicle to a signal light in a closed area according to claim 5, wherein: After the two-dimensional verification, the performance evaluation engine is deployed. Specifically, based on the data stream after time stamp alignment, the mean and variance of the delay from signal light recognition to vehicle control response are calculated. When the delay exceeds 150ms or the variance is greater than 30ms, the performance evaluation engine is used to calculate the delay from signal light recognition to vehicle control response. 2 When the test fails, the test is automatically marked as invalid and an alarm signal is triggered to the monitoring platform.

Images