A green wave effect inspection method and system based on a unmanned aerial vehicle

Abstract

This invention discloses a method and system for inspecting green wave effects based on unmanned aerial vehicles (UAVs), relating to the cross-technical field of traffic signal control effect evaluation and UAV applications. The method includes the following steps: S1, acquiring a green wave scheme and generating a UAV flight path extending along a continuous road segment; S2, the UAV autonomously flies along the flight path, simultaneously acquiring continuous images and identifying target convoys in real time, obtaining virtual parking data and target parking data; S3, fusing the virtual parking data and target parking data at the terminal side, first eliminating non-signal factors, and obtaining speed deviation, abnormal parking ratio, continuous parking degree, directional parking ratio, and queue clearing delay indicators; S4, uploading abnormal events where the indicators exceed the corresponding thresholds in real time, triggering green wave timing optimization, implementing full coverage of the continuous road segment where the green wave is located through the UAV positioning trajectory, and achieving blind-spot-free acquisition of the green wave operation status by replacing the fixed ground perspective with an aerial moving perspective.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of traffic signal control effect evaluation and drone application, specifically to a drone-based green wave effect inspection method and system. Background Technology

[0002] Green wave coordinated control on main roads, using a unified timing method at continuous intersections to allow convoys to pass through without stopping at a set speed, is a core means of improving the traffic efficiency of urban arterial roads. However, whether the green wave bandwidth can maintain a constant match with the actual traffic flow speed and arrival patterns remains a long-standing pain point in the industry, lacking scalable and closed-loop verification methods. Existing technical approaches can be categorized into two types: one is "fixed detectors + background models," which involves deploying coils, videos, or radars at intersection cross-sections. This only obtains discrete vehicle pulses and cannot capture key processes such as speed decay, convoy dispersion, and secondary stops within the road segment, creating a structural data gap of "accurate cross-sections but blind road segments." The other is "human-driven test vehicles," which drive at green wave speeds and record the number of stops. This method is constrained by manpower, safety, and traffic control, and can only be implemented at low frequencies. Furthermore, the recorded results are highly subjective and almost impossible to conduct during peak hours, resulting in data gaps during the periods when data is most needed. Both methods together contribute to the situation of "low spatiotemporal resolution and difficulty in identifying causality" in green wave effect evaluation, leading to a long-term reliance on trial and error for signal optimization.

[0003] More problematic is that parking caused by non-signal factors such as traffic accidents, road construction, and illegal parking is highly coupled with traditional "inappropriate timing." Existing systems lack front-end identification capabilities and can only report all parking events, generating numerous invalid work orders daily. This results in a significant waste of operational resources without addressing the true timing defects. This misjudgment problem has persisted since the inception of green wave technology, and the industry has consistently failed to propose a systematic solution that balances "continuous perception throughout the process" with "automatic interference removal," severely hindering the refined and intelligent evolution of green wave coordination.

[0004] For example, there is a Chinese patent with publication number CN104699956B, which relates to a method for evaluating the coordination effect of green wave on main roads based on mobile terminals. It uses mobile phone GPS to draw a spatiotemporal map of a single vehicle trajectory and outputs indicators such as the number of stops and travel time. However, the Chinese patent with publication number CN104699956B can only obtain single-point mobile data on the ground, and still cannot cover the entire road section between intersections. It also does not identify and eliminate interference events such as accidents, construction, and illegal parking. The evaluation results are easily affected by non-signal parking, making it difficult to achieve continuous and accurate collection of the green wave operation status. Summary of the Invention

[0005] To address the issues of incomplete green wave status due to inaccurate cross-section detection in existing roadside detection systems, as well as the maintenance burden caused by non-signal factors being misjudged as timing anomalies, this invention proposes a green wave effect inspection method and system based on unmanned aerial vehicles (UAVs). The UAV achieves full coverage of the continuous road sections where the green wave is located through centimeter-level positioning trajectory, and replaces the fixed ground perspective with an aerial moving perspective to achieve blind-spot-free acquisition of the green wave operation status.

[0006] A further objective of this invention is to identify and eliminate accidents, construction, and illegal parking incidents at the end-point by relying on an airborne lightweight visual model, while retaining only the uploaded data of real timing anomalies, thereby achieving an accurate evaluation of the green wave operation status with a false alarm rate approaching zero.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a green wave effect inspection method based on unmanned aerial vehicles (UAVs), comprising the following steps: S1, obtain the green wave band scheme and generate the drone flight path extending along the continuous road segment; S2, the drone flies autonomously along the route, simultaneously collecting continuous images and identifying the target convoy in real time, obtaining virtual parking data and target parking data; S3 integrates virtual parking data and target parking data on the terminal side. After eliminating non-signal factors, it obtains speed deviation, abnormal parking ratio, continuous parking degree, directional parking ratio and queue clearing delay indicators. S4 uploads abnormal events where the indicator exceeds the corresponding threshold in real time, triggering green wave timing optimization.

[0008] In this technical solution, the drone continuously covers the entire green wave band with centimeter-level trajectory. After the lightweight AI on the drone eliminates non-signal parking in real time, it calculates five indicators: "speed deviation, abnormal parking, continuous parking, proportion of parking in different directions, and queue not cleared". It outputs timing abnormal alarms to achieve green wave effect with no blind spots, zero false alarms, low cost, and high frequency closed-loop optimization.

[0009] Preferably, in step S3, the elimination of non-signal factors includes: running a lightweight vision model locally on the UAV to simultaneously detect traffic accidents, road construction, and illegal parking; if the above events are detected, they are marked as external interference and the abnormal judgment of the green wave strategy at the corresponding intersection is skipped, so as to avoid accidental events such as accidents and construction being misjudged as timing defects and to greatly reduce ineffective operation and maintenance.

[0010] Preferably, in step S3, obtaining the abnormal parking ratio index includes: performing confidence-weighted fusion of the simulated parking ratio of the drone and the actual parking ratio of the fleet. When the fusion value is greater than 20% and external interference is eliminated, the intersection is determined to be an abnormal parking point. Dual-source weighting cancels out single-source errors, making the determination of whether the timing itself is reasonable more robust.

[0011] Preferably, the ratio of the number of times the drone simulates parking to the total number of times it passes through within the same inspection cycle is assigned a weight of 0.5, and the ratio of the actual number of parking vehicles in the image recognition to the total number of captured vehicles is assigned a weight of 0.5. The two weighted results are added together to obtain the abnormal parking ratio. A confidence level is introduced to weight and fuse the two, so that the two types of error sources cancel each other out, and an estimate that is closer to "whether the timing itself is reasonable" is obtained. The confidence level can be flexibly adjusted to adapt to different traffic scenarios.

[0012] Preferably, in step S3, the continuous parking index is obtained as follows: when the proportion of abnormal parking at three consecutive intersections is greater than 20% and external interference has been eliminated, it is determined to be a continuous parking event, and it is suggested that the green light timing between intersections needs to be optimized, expanding from "point" abnormality to "line" imbalance, quickly locating the road segments that need to be coordinated and cascaded to correct, and avoiding blindly adjusting parameters at each intersection.

[0013] Preferably, the calculation steps for the abnormal parking percentage at three consecutive intersections include: taking the current intersection as the center, extending forward and backward by one intersection to form a three-intersection window; calculating the abnormal parking percentage at each intersection; if the abnormal parking percentage at all three intersections within the window is >20%, a continuous parking alarm is immediately generated. The sliding window mechanism ensures spatial continuity and prevents sporadic single-point parking from being amplified into continuous imbalance.

[0014] Preferably, in step S3, obtaining the directional parking ratio index includes: separately calculating the proportion of intersections judged as abnormal parking in the up and down directions to the total number of intersections in that direction. When the proportion is greater than 1 / 3 and external interference has been eliminated, it is determined that the total number of directional parking times is too high, and it is suggested that the green wave timing parameters of the corresponding direction need to be adjusted as a whole. Macro statistics reveal the systematic imbalance of the entire path and guide the overall correction of the common cycle or design speed.

[0015] Preferably, in step S3, obtaining the queue clearing delay index includes: comparing the difference between the start time of the green light bandwidth and the actual time when the vehicle at the front of the queue, tracked by the drone, crosses the line; simultaneously, calculating the equivalent length of vehicles remaining at the moment the green light turns on; when the delay is greater than 10 seconds and the remaining length is greater than 35 meters and external interference has been eliminated, it is determined that the queue has not been cleared, and a prompt is made to advance the green light start time, accurately identifying the scenario where the tail of the previous cycle occupies the bandwidth of the current cycle, and avoiding the interruption of the green wave caused by the green wave convoy stopping twice.

[0016] This invention also employs the following technical solution: a drone-based green wave effect inspection system, implementing the aforementioned drone-based green wave effect inspection method, comprising: a positioning module locking the drone's flight trajectory onto the centerline of the target lane in the green wave zone; a range perception module acquiring images; a calculation module identifying vehicles and traffic events in real time based on the images, calculating speed deviation, abnormal parking percentage, continuous parking degree, directional parking percentage, and queue clearing delay indicators, completing the elimination of non-signal factors and the determination of green wave anomalies; and a communication module uploading the anomaly determination results to the traffic control center to trigger timing optimization.

[0017] Preferably, the computing module runs lightweight visual models, including the YOLOv5-Lite traffic accident detection model, the ResNet18 road construction detection model, and the illegal parking and lane occupation detection model based on trajectory static discrimination. When at least one lightweight visual model outputs an event, the corresponding intersection is marked as external interference and the green wave anomaly judgment is skipped. The mature lightweight architecture is used to achieve millisecond-level event screening on the device side, ensuring that the index calculation focuses only on the signal itself and greatly improving the reliability of the alarm.

[0018] The beneficial effects of this invention are: 1) Reduce resource dependence and improve inspection efficiency: No need to invest in manpower and ground vehicles, a single drone can complete the inspection of multiple routes within 1 hour; and the flight speed matches the green wave range, avoiding the speed deviation of manual operation. 2) Detailed quantitative assessment: By using queue modeling to distinguish the passability differences of the leading, middle, and trailing vehicles, specific directions are provided for timing optimization, avoiding subjective human bias; 3) Improve the accuracy and effectiveness of evaluation results by excluding non-signal-based parking events. Attached Figure Description

[0019] Figure 1 This is a flowchart of a green wave effect inspection method based on drones according to the present invention.

[0020] Figure 2 This is a flowchart of steps S3 and S4 of a green wave effect inspection method based on drones according to the present invention. Detailed Implementation

[0021] Example 1 This embodiment provides a method for inspecting green wave effects based on unmanned aerial vehicles (UAVs), such as... Figure 1 As shown, it includes the following steps.

[0022] Step S1: Green wave scheme acquisition and path planning.

[0023] Obtain the list of intersections, road segments, green wave timing and scheme for the route to be evaluated from the traffic control center, including the name of each intersection, green light start time, green light duration, green wave design speed, and estimated stop light parameters.

[0024] The flight path is generated based on the latitude and longitude of the green wave route to be evaluated, including both uphill and downhill green wave routes, with the route maintained in the middle lane as much as possible. The starting point is 1 kilometer before the first intersection of the route, and the ending point is 1 kilometer after the last intersection, to ensure effective tracking and coordination of the maximum vehicle queue in the direction when the inspection task begins.

[0025] Step S2: In-flight data acquisition and real-time processing.

[0026] After taking off along a preset route and at a preset speed, the drone performs three tasks during flight: collecting simulated driving behavior data, collecting actual on-road vehicle targets, and identifying unexpected factors.

[0027] During flight, the drone collects images at a frequency of 10Hz, identifies the position of all vehicles within its field of view through the images, and maintains its own position in the middle of the vehicle queue. When in the entrance area, it remains in the middle of the queue length of the lane in the coordinated direction.

[0028] The local computing module calculates the time it takes for the drone to reach the stop line at each intersection based on real-time positioning data, and compares it with the real-time green light sequence: if the arrival time is within the range of the green light start time + green light duration, it is recorded as passing through the green light; if it is not within the range, it is recorded as a virtual stop. At the same time, the remaining red light waiting time is calculated. The passing result of each intersection must be recorded throughout the flight.

[0029] Furthermore, for the image data during the inspection process, the system compares the front and rear images to correlate the vehicle's front and rear trajectories, determines whether the vehicle is stopped (speed less than 7km / h and duration greater than 5 seconds), whether it is in the driving state of the road segment, and whether it is located in the coordinated direction lane at the entrance segment. It also records data such as the time each vehicle takes to pass the stop line at each intersection, its driving speed in the road segment, the number of stops, and the duration of stops.

[0030] Meanwhile, the local computing module runs multiple small models for detecting and recognizing unexpected events in real time, including traffic accident detection models, road construction recognition models, and vehicle illegal parking and lane occupation recognition models, to identify unexpected situations that may affect the effectiveness of green wave traffic.

[0031] Step S3: Anomaly identification and platform alarm.

[0032] Before running the event evaluation model, the UAV's local computing module first starts the unexpected event detection model. After preprocessing the collected image data (such as grayscale conversion and noise reduction), it inputs it into multiple lightweight recognition models to prioritize the elimination of interfering factors. The flowcharts for steps S3 and S4 are shown below. Figure 2 The specific model design will be described below.

[0033] Traffic accident detection model: The YOLOv5-Lite lightweight target detection algorithm is adopted. The training set contains traffic accident feature samples such as vehicle collisions, rollovers, and people falling. The model input is a 4K high-definition image from a drone, and the output is whether a traffic accident exists and the coordinates of the event location.

[0034] Road construction detection model: Based on the ResNet18 lightweight classification network, it focuses on identifying traffic cones (orange cylindrical targets), construction warning signs (red background with white text / yellow background with black text), and construction vehicles (trucks with engineering markings) in the construction area. The model improves the robustness of recognition through multi-feature fusion (color, shape, texture).

[0035] Illegal parking detection model: Combining vehicle motion trajectory analysis and location determination, the model uses continuous images from drones to determine whether a vehicle is stationary. If a stationary vehicle is located in a non-parking area for a duration of ≥20 seconds, it is determined to be illegally parked and occupying the road.

[0036] When the unexpected event detection model outputs the existence of the aforementioned events, the event is marked as a "non-green wave scheme problem," and the event type, location, and duration are recorded separately. The green wave strategy anomaly determination for the corresponding intersection in the event evaluation model is skipped. If no unexpected event is detected, the event evaluation model is activated, and the presence of green wave strategy-related anomalies is determined sequentially according to the indicators in step S4. Data is uploaded, including: anomaly type, intersection where the anomaly occurred, anomaly time period, judgment basis data (such as speed deviation rate, parking percentage, queue length, etc.), and on-site video footage. After receiving the data, the traffic control platform classifies, stores, and visualizes the data according to the problem type, such as marking the anomaly location on an electronic map and using different colors to distinguish the anomaly type: red represents a green wave scheme problem, yellow represents an unexpected event, and a tiered alarm mechanism is triggered simultaneously.

[0037] Step S4: Based on the simulated driving data, target vehicle dynamic data and intersection traffic status data collected by the drone, construct multi-dimensional evaluation indicators to quantify the abnormality of green wave coordination effect.

[0038] The calculation and judgment logic of each indicator will be explained in detail below.

[0039] First, evaluate the rationality of the actual speed versus the preset speed. Based on the speed deviation between the actual average speed of the road segment and the target driving speed of the route to be inspected, determine the suitability of the speed parameters with the actual traffic flow.

[0040] The preset design speed of the green wave route is defined as v0 (unit: km / h). This parameter is extracted from the green wave scheme obtained from the traffic control center and corresponds to the target driving speed of the route to be inspected.

[0041] The actual average speed of a road segment is defined as vr. This parameter is obtained by the UAV during flight, which tracks and coordinates the target convoy (a group of vehicles consisting of the lead, middle, and tail vehicles) using image recognition. The arithmetic mean of the speeds of each vehicle within the convoy is calculated, representing the ratio of the sum of the speeds of individual vehicles to the total number of vehicles in the target convoy. The speed of an individual vehicle is calculated based on the changes in vehicle position and time differences between UAV image frames. To quantify the degree of speed deviation, a speed deviation rate S is introduced, which is obtained as the ratio of the absolute value of the difference between the actual average speed of the road segment and the preset design speed of the green wave route to the preset design speed of the green wave route.

[0042] When S>20%, it is determined that the actual speed is unreasonable compared with the preset speed, indicating that the speed parameters in the green wave scheme are not well adapted to the actual traffic flow.

[0043] Second, evaluation of abnormal parking spots.

[0044] The determination of abnormal parking points requires a combination of data from both drone-simulated parking events and the actual parking situation of the target vehicle fleet to avoid misjudgment due to bias in a single data point.

[0045] In the dimension of drone-simulated parking events, the drone simulates vehicle movement at green wave speed. If the drone reaches the stop line outside the green light window, a "virtual stop" is recorded. This dimension reflects the degree of matching between signal timing and ideal speed, but it cannot reflect whether "real traffic also stops".

[0046] In terms of the actual parking situation of the target fleet, the number of vehicles actually stopped and waiting in the target fleet is counted using an airborne visual model. This dimension reflects the actual state of traffic flow after being affected by factors such as signals, queues, and interference, but may include non-signal parking such as accidents, construction, and illegal parking.

[0047] Data relying solely on drone-simulated parking events will overestimate "timing errors," while data relying solely on the actual parking situation of the target fleet will amplify "non-signal interference." Therefore, confidence levels are introduced to weight and fuse the two, allowing the two error sources to cancel each other out, resulting in an estimate that more closely approximates whether "timing itself is reasonable."

[0048] In this embodiment, the average parking percentage Di at the i-th intersection is calculated based on the number of virtual stops recorded by the UAV when it passes through intersection i during the inspection cycle because the time to reach the stop line is not during the green light period, the total number of times the UAV passes through intersection i during the inspection cycle, the number of vehicles in the target convoy identified by the UAV image that are waiting at intersection i, the total number of vehicles in the target convoy captured by the UAV at the intersection, and the confidence weight.

[0049] Specifically, the average parking percentage at the i-th intersection is obtained by adding the simulated parking items and the actual parking items at that intersection.

[0050] The simulated parking item is obtained by multiplying the confidence weight by the ratio of the number of virtual parkings recorded by the drone when it passes through intersection i during the inspection cycle because the time it reaches the stop line is not during the green light period, to the total number of times the drone passes through intersection i during the same inspection cycle.

[0051] The actual parking item is obtained by multiplying the ratio of the number of vehicles waiting at intersection i in the target convoy identified by the drone image to the total number of vehicles in the target convoy captured by the drone at that intersection by the complement of the confidence weight.

[0052] The overall parking percentage is obtained by linearly balancing the "virtual parking percentage" and the "actual parking percentage" using confidence weights. The confidence weights can be configured by the user and are typically 0.5.

[0053] When Di>20% and subsequent accidental event detection models confirm that there are no interfering factors such as traffic accidents, road construction, or illegal parking, intersection i is determined to be an abnormal parking point, indicating that there may be a deviation in the green wave timing at this intersection, and the green light duration of the coordinated phase needs to be optimized.

[0054] In this technical solution, threshold discrimination and causal filtering are combined. When Di > 20%, the parking phenomenon is considered to be significantly higher than the tolerance for random fluctuations and enters the candidate set of "suspected abnormal parking spots". Only when the subsequent lightweight visual model confirms that there are no external events such as traffic accidents, road construction, or illegal parking at the intersection will it be finally determined as an "abnormal parking spot", thereby eliminating "non-signal parking" and ensuring that the conclusion points to timing deviation.

[0055] Third, continuous parking evaluation.

[0056] The determination of continuous parking is based on the spatial continuity of parking status at intersections. Three consecutive intersections on the green wave line are selected, and the actual parking percentages Di, Di+1, and Di+2 at each of the three intersections are calculated. When all three percentages are greater than 20% and interference from unexpected events is excluded, it is determined to be continuous parking, indicating that the timing coordination of this continuous segment in the green wave line is insufficient, and the green light sequence between the consecutive intersections needs to be optimized.

[0057] Fourth, evaluation of total number of parking sessions.

[0058] Green wave traffic is a "line" control concept. An anomaly at a single intersection is insufficient to indicate a failure of the scheme. A "directional" macro-level indicator is needed to measure the overall coordination level. Therefore, the parking intersection ratio is introduced, which is the ratio of the number of intersections judged as abnormal parking on a green wave path to the total number of intersections, as a statistical characteristic of directional failure.

[0059] The percentage of parking intersections in both directions is calculated based on the number of parking intersections CSu in the up direction, the total number of parking intersections Cu in the up direction, the number of parking intersections CSd in the down direction, and the total number of parking intersections Cd in the down direction.

[0060] Specifically, the percentage of intersections with stops on the uphill direction is calculated as the ratio of the number of intersections with stops on the uphill direction (CSu) to the total number of intersections on the uphill direction (Cu); the percentage of intersections with stops on the downhill direction is calculated as the ratio of the number of intersections with stops on the downhill direction (CSd) to the total number of intersections on the downhill direction (Cd). Here, the number of intersections with stops on the uphill direction (CSu) is the total number of intersections on the uphill direction where Di > 20%, and the same applies to the downhill direction.

[0061] When ru or rd > 1 / 3 and interference from unexpected events is excluded, it is determined that the total number of stops is too high, and the timing parameters of the green wave scheme for that direction need to be adjusted as a whole.

[0062] The evaluation of total number of stops also incorporates threshold discrimination and causal filtering. When ru or rd > 1 / 3, it is considered that more than 1 / 3 of the intersections in that direction simultaneously experience timing-related stops, and the probability is no longer a random fluctuation but a systematic imbalance. Only intersections where "accidents, construction, and illegal parking have been ruled out" are retained to ensure that the increase in the proportion truly stems from signal timing rather than external emergencies.

[0063] Fifth, the queue of reviews was not cleared.

[0064] If the queue from the previous cycle has not completely dissipated when the green light turns on, the coordinated phase bandwidth is occupied. Subsequent green wave convoys will encounter "residual" vehicles, causing a secondary stop and resulting in a "wave break" in the green wave. Therefore, it is necessary to simultaneously measure both "time" and "space" dimensions. In the time dimension, the duration of the queue remaining after the green light has turned on is used to determine whether queue clearing lags behind the start of the green light window. In the spatial dimension, the physical queue length left from the previous cycle is used to determine whether the scale of the remaining queue has reached a level that affects bandwidth.

[0065] By combining the green light sequence with the dynamics of the queuing vehicles, the bandwidth start time tg of the coordinated phase i at the intersection is obtained through real-time traffic light information, which is the expected time for the first vehicle to cross the stop line. The drone tracks the frontmost vehicle in the queue through continuous image frames and records the time when its front wheels cross the stop line to obtain the time tp when the first vehicle crosses the stop line. Then, the queue clearing time Tq is the difference between the time when the first vehicle crosses the stop line and the bandwidth start time of the coordinated phase i at the intersection.

[0066] Meanwhile, the queue length Lq is defined. The number of vehicles in the lane of the coordinated direction at intersection i is calculated by the drone through image recognition of the number of vehicles at the moment the green light turns on, combined with the average lane occupancy length of vehicles on urban roads (taken as 7m / vehicle).

[0067] When Tq > 10 seconds and Lq > 35 meters, and interference from unexpected events is excluded, it is determined that the queue has not been cleared, indicating that the queue of vehicles in the previous cycle at this intersection has not completely dissipated, and the green light start time needs to be optimized.

[0068] Example 2 This embodiment provides a green wave effect inspection system, configured in a drone, which meets the requirements of high-precision positioning, range awareness, low-latency communication, and local real-time computing. It includes a positioning module, a range awareness module, a communication module, and a computing module.

[0069] The positioning module locks the drone's flight path onto the center line of the target lane in the green wave zone.

[0070] In this embodiment, the positioning module adopts BeiDou / GPS dual-mode positioning + RTK real-time differential technology, and the positioning accuracy needs to reach the centimeter level, with a horizontal error of ≤0.5 meters and an elevation error of ≤1 meter. This accuracy can ensure that the deviation between the UAV flight trajectory and the center line of the target lane of the green wave route does not exceed 1 meter, resulting in a higher path matching degree when simulating vehicle driving and avoiding evaluation errors caused by trajectory deviation.

[0071] The range perception module acquires images. In this embodiment, the range perception module is equipped with a 4K high-definition wide-angle camera. The camera must have a horizontal field of view of more than 120° and a resolution of not less than 3840×2160. It is equipped with a night infrared fill light function to ensure that the vehicle outline and parking status can be clearly identified both day and night.

[0072] The calculation module identifies vehicles and traffic events in real time based on images, calculates speed deviation, abnormal parking ratio, continuous parking degree, directional parking ratio, and queue clearing delay index, and completes the elimination of non-signal factors and the determination of green wave anomalies.

[0073] In this embodiment, the computing module is equipped with a quad-core processor and at least 8GB of memory, supporting the local execution of lightweight data processing algorithms (such as vehicle detection and parking recognition), eliminating the need to transmit all raw data back to the control center, reducing communication bandwidth usage, and improving analysis efficiency.

[0074] When the computing module runs, it can directly execute the UAV-based green wave effect inspection method described in Example 1.

[0075] The lightweight visual models include the YOLOv5-Lite traffic accident detection model, the ResNet18 road construction detection model, and the illegal parking and lane occupation detection model based on trajectory static discrimination. When at least one lightweight visual model outputs an event, the corresponding intersection is marked as external interference and the green wave anomaly judgment is skipped.

[0076] The communication module uploads the anomaly detection results to the traffic control center, triggering timing optimization.

[0077] In this embodiment, the communication module uses 4G / 5G communication, and the communication latency needs to be ≤50ms to ensure that the UAV can receive the green wave scheme update issued by the control center in real time and quickly upload alarm information.

[0078] In addition, the drone must have a constant speed cruise function with a speed adjustment range of 0-80km / h, matching the green wave design speed of different roads to ensure the stability of simulated driving behavior.

[0079] The workflow of a drone-based green wave effect inspection system in this embodiment is described in detail below.

[0080] The first step is to preset the parameters before the green wave inspection.

[0081] Green wave scheme acquisition: Obtain the list of intersections, road segments, green wave timing and scheme of the route to be evaluated from the traffic control center, including the name of each intersection, green light start time, green light duration, green wave design speed, and estimated stop light parameters.

[0082] Path planning: Generate flight routes based on the latitude and longitude of the green wave route to be evaluated, including both uphill and downhill green wave routes, keeping the route as close to the middle lane as possible. The starting point is 1 kilometer before the first intersection and the ending point is 1 kilometer after the last intersection to ensure effective tracking and coordination of the maximum vehicle queue when the inspection task begins.

[0083] Secondly, there is in-flight data acquisition and real-time processing.

[0084] After taking off along a preset route and at a preset speed, the drone performs three tasks during flight: collecting simulated driving behavior data, collecting actual on-road vehicle targets, and identifying unexpected factors.

[0085] During flight, the drone collects images at a frequency of 10Hz, identifies the position of all vehicles within its field of view through the images, and maintains its own position in the middle of the vehicle queue. When in the entrance area, it remains in the middle of the queue length of the lane in the coordinated direction.

[0086] The local computing module calculates the time it takes for the drone to reach the stop line at each intersection based on real-time positioning data, and compares it with the real-time green light sequence: if the arrival time is within the range of the green light start time + green light duration, it is recorded as passing through the green light; if it is not within the range, it is recorded as a virtual stop. At the same time, the remaining red light waiting time is calculated. The passing result of each intersection must be recorded throughout the flight.

[0087] Furthermore, for the image data during the inspection process, the system compares the front and rear images to correlate the vehicle's front and rear trajectories, determines whether the vehicle is stopped (speed less than 7km / h and duration greater than 5 seconds), whether it is in the driving state of the road segment, and whether it is located in the coordinated direction lane at the entrance segment. It also records data such as the time each vehicle takes to pass the stop line at each intersection, its driving speed in the road segment, the number of stops, and the duration of stops.

[0088] Meanwhile, the local computing module runs multiple small models for detecting and recognizing unexpected events in real time, including traffic accident detection models, road construction recognition models, and vehicle illegal parking and lane occupation recognition models, to identify unexpected situations that may affect the effectiveness of green wave traffic.

[0089] The next is the event evaluation model.

[0090] The core of the event evaluation model is to construct multi-dimensional evaluation indicators based on simulated driving data, target vehicle dynamic data, and intersection traffic status data collected by drones, and to quantify the abnormality of green wave coordination effect, including the following evaluation indicators.

[0091] 1. Evaluation of the Reasonableness of Actual Speed ​​vs. Preset Speed: The preset design speed of the green wave route is defined as v0 (unit: km / h). This parameter is extracted from the green wave scheme obtained from the traffic control center and corresponds to the target driving speed of the route to be inspected. The actual average speed of the road segment is defined as vr. This parameter is obtained by UAVs tracking and coordinating the target convoy (a group of vehicles consisting of the lead, middle, and tail vehicles) during flight, collecting the arithmetic mean of the driving speed of each vehicle in the convoy within the road segment. To quantify the degree of speed deviation, a speed deviation rate S is introduced. When S > 20%, it is determined that the actual speed is unreasonable compared to the preset speed, indicating that the speed parameters in the green wave scheme are not well adapted to the actual traffic flow.

[0092] 2. Evaluation of Abnormal Parking Spots: The determination of abnormal parking spots requires the combination of two-dimensional data: simulated parking events by drones and the actual parking situation of the target fleet. This avoids misjudgment due to bias in a single data point. The average parking percentage Di at the i-th intersection is calculated. When Di > 20% and subsequent accidental event detection models confirm that there are no interfering factors such as traffic accidents, road construction, or illegal parking, intersection i is determined to be an abnormal parking spot. This indicates that there may be a deviation in the green wave timing at this intersection, and the green light duration of the phase needs to be optimized and coordinated.

[0093] 3. Continuous Parking Evaluation: The determination of continuous parking is based on the spatial continuity of parking status at intersections. Three consecutive intersections on the green wave route are selected, and the actual parking percentages Di, Di+1, and Di+2 at each of the three intersections are calculated. When all three percentages are greater than 20% and interference from unexpected events is excluded, it is determined to be continuous parking, indicating insufficient timing coordination of the continuous road segment in the green wave route, and the green light timing sequence between the consecutive intersections needs to be optimized.

[0094] 4. Evaluation of total number of stops: The percentage of stops in both directions is calculated based on the number of stops in the up direction (the total number of stops with Di>20% in the up direction), the total number of stops in the up direction, the number of stops in the down direction, and the total number of stops in the down direction. When the percentage of stops in either direction is greater than 1 / 3 and interference from unexpected events is excluded, it is determined that the total number of stops is too high, and the timing parameters of the green wave scheme for that direction need to be adjusted as a whole.

[0095] 5. Evaluation of Uncleared Queues: Combining the green light sequence with the dynamics of the queued vehicles, the bandwidth start time of the i-coordinated phase at the intersection (i.e., the expected time for the first vehicle to cross the stop line) is obtained through real-time traffic light information. The drone tracks the frontmost vehicle in the queue through continuous image frames, records the time when its front wheels cross the stop line, obtains the time when the first vehicle crosses the stop line, and calculates the difference to obtain the queue clearing time.

[0096] Meanwhile, based on the queue length, the number of vehicles in the coordinated direction lane at intersection i is calculated by drone through image recognition at the moment the green light turns on, combined with the average lane occupancy length of vehicles on urban roads (taken as 7m / vehicle).

[0097] If the queue clearing time is greater than 10 seconds and the queue length is greater than 35 meters, and interference from unexpected events is excluded, the queue is judged as not cleared, indicating that the queue of vehicles in the previous cycle at this intersection has not completely dissipated, and the green light start time needs to be optimized.

[0098] Finally, there is the anomaly detection and platform alarm mechanism.

[0099] Before running the event evaluation model, the UAV local computing module first starts the unexpected event detection model. After preprocessing the collected image data (such as grayscale conversion and noise reduction), it inputs it into multiple lightweight recognition models to achieve priority elimination of interference factors. The specific model design is as follows.

[0100] Traffic accident detection model: The YOLOv5-Lite lightweight target detection algorithm is adopted. The training set contains traffic accident feature samples such as vehicle collisions, rollovers, and people falling. The model input is a 4K high-definition image from a drone, and the output is whether a traffic accident exists and the coordinates of the event location.

[0101] Road construction detection model: Based on the ResNet18 lightweight classification network, it focuses on identifying traffic cones (orange cylindrical targets), construction warning signs (red background with white lettering / yellow background with black lettering) and construction vehicles (trucks with engineering markings) in the construction area. The model improves the robustness of recognition through multi-feature fusion (color, shape, texture).

[0102] Illegal parking detection model: Combining vehicle motion trajectory analysis and location determination, the model uses continuous images from drones to determine whether a vehicle is stationary. If a stationary vehicle is located in a non-parking area for a duration of ≥20 seconds, it is determined to be illegally parked and occupying the road.

[0103] When the accidental event detection model outputs the existence of the above-mentioned events, the event is marked as a "non-green wave scheme problem", the event type, location, and duration are recorded separately, and the abnormal determination of the green wave strategy at the corresponding intersection in the event evaluation model is skipped.

[0104] If no unexpected event is detected, the event evaluation model is activated to determine whether there are any anomalies related to the green wave strategy according to the above indicators, and the data uploaded includes: anomaly type, intersection where the anomaly occurred, time period of the anomaly, data on the basis for judgment (such as speed deviation rate, parking ratio, queue length, etc.), and on-site video clips.

[0105] After receiving the data, the traffic control platform classifies, stores, and visualizes the data according to the type of problem. For example, it marks abnormal locations on an electronic map and uses different colors to distinguish the types of abnormalities: red represents problems with the green wave scheme, yellow represents unexpected events, and triggers a graded alarm mechanism.

Claims

1. A method for inspecting green wave effect based on a UAV, characterized in that, Includes the following steps: S1, obtain the green wave band scheme and generate the drone flight path extending along the continuous road segment; S2, the drone flies autonomously along the route, simultaneously collecting continuous images and identifying the target convoy in real time, obtaining virtual parking data and target parking data; S3 integrates virtual parking data and target parking data on the terminal side. After eliminating non-signal factors, it obtains speed deviation, abnormal parking ratio, continuous parking degree, directional parking ratio and queue clearing delay indicators. S4 uploads abnormal events where the indicator exceeds the corresponding threshold in real time, triggering green wave timing optimization. 2.The UAV-based green wave effect inspection method of claim 1, wherein, In step S3, the elimination of non-signal factors includes: running a lightweight vision model locally on the drone to simultaneously detect traffic accidents, road construction, and illegal parking; if the above events are detected, they are marked as external interference and the abnormal determination of the green wave strategy at the corresponding intersection is skipped.

3. The method for inspecting green wave effects based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, In step S3, the abnormal parking ratio index is obtained by: performing a confidence-weighted fusion of the simulated parking ratio of the drone and the actual parking ratio of the fleet. When the fusion value is greater than 20% and external interference is eliminated, the intersection is determined to be an abnormal parking point.

4. The method for inspecting green wave effects based on unmanned aerial vehicles (UAVs) according to claim 3, characterized in that, The ratio of the number of simulated parking times to the total number of times the drone passes through within the same inspection cycle is assigned a weight of 0.5, and the ratio of the actual number of parking times of the fleet recognized by image recognition to the total number of captured vehicles is assigned a weight of 0.

5. The two weighted results are added together to obtain the abnormal parking ratio.

5. A method for inspecting green wave effects based on unmanned aerial vehicles (UAVs) according to claim 1, 3, or 4, characterized in that, In step S3, the continuous parking index is obtained as follows: when the proportion of abnormal parking at three consecutive intersections is greater than 20% and external interference has been eliminated, it is determined to be a continuous parking event, and it is suggested that the green light timing between intersections needs to be optimized.

6. The method for inspecting green wave effects based on unmanned aerial vehicles (UAVs) according to claim 5, characterized in that, The steps for calculating the percentage of abnormal parking at the three consecutive intersections include: taking the current intersection as the center, extending forward and backward by one intersection to form a three-intersection window; calculating the percentage of abnormal parking at each intersection; if the percentage of abnormal parking at all three intersections within the window is >20%, a continuous parking alarm is immediately generated.

7. The method for inspecting green wave effects based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, In step S3, the directional parking ratio indicator is obtained by: calculating the proportion of intersections judged as abnormal parking in the up and down directions to the total number of intersections in that direction. When the proportion is greater than 1 / 3 and external interference has been eliminated, it is determined that the total number of directional parking times is too high, and the corresponding green wave timing parameters need to be adjusted as a whole.

8. The method for inspecting green wave effects based on unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, In step S3, the queue clearing delay index is obtained by comparing the difference between the start time of the green light bandwidth and the actual time when the vehicle at the front of the queue crossed the line as tracked by the drone, and simultaneously calculating the equivalent length of the vehicles that were stuck at the moment the green light turned on. When the delay is greater than 10 seconds and the stuck length is greater than 35 meters and external interference has been eliminated, it is determined that the queue has not been cleared, and a prompt is made to advance the green light start time.

9. A drone-based green wave effect inspection system, implementing the drone-based green wave effect inspection method according to any one of claims 1-8, characterized in that, include: The positioning module locks the drone's flight path onto the center line of the target lane in the green wave band; The range sensing module acquires images; The calculation module identifies vehicles and traffic events in real time based on images, calculates speed deviation, abnormal parking ratio, continuous parking degree, directional parking ratio, and queue clearing delay index, and completes non-signal factor elimination and green wave anomaly determination; the communication module uploads the anomaly determination results to the traffic control center to trigger timing optimization.

10. A UAV-based green wave effect inspection system according to claim 9, characterized in that, The computing module runs lightweight visual models, including the YOLOv5-Lite traffic accident detection model, the ResNet18 road construction detection model, and the illegal parking and lane occupation detection model based on trajectory static discrimination. When at least one lightweight visual model outputs an event, the corresponding intersection is marked as external interference and the green wave anomaly judgment is skipped.

Images