A road construction intelligent monitoring method based on unmanned aerial vehicle video stream

Abstract

The application discloses a kind of road construction intelligent monitoring methods based on unmanned aerial vehicle video stream, it is related to video identification technical field, including after obtaining video stream detection and track construction target and select main road, according to trigger condition start construction session;In session, target is filtered with space and confidence, calculates its road width normalization position, and with target identity identification as main key cross frame accumulates time sequence record;When construction facility is not detected continuously over threshold, end session, generate road width position and time joint two-dimensional semantic graph and layout graph by record summary;For cone bucket, first in road width direction, density peak clustering and keeping representative frame unchanged Gaussian smoothing are carried out, then in road width time two-dimensional space, secondary clustering smoothing is carried out to suppress jitter;Layout is arranged to layout graph, generate statistical data and pack spatiotemporal metadata, carry out session level single report. Panoramic trace and session level closed loop output of construction situation are realized, solve fragmentation and jitter problem of frame-by-frame monitoring.

Description

Technical Field

[0001] This invention relates to the field of video recognition technology, specifically to an intelligent monitoring method for road construction based on UAV video streams. Background Technology

[0002] In the daily maintenance of operating highways, the supervision of construction safety in scenarios such as road occupancy and road-related operations is a core aspect of ensuring road traffic safety and standardized construction practices. With the rapid development of drone technology, using drones equipped with high-definition cameras for aerial inspections of construction sections has become an efficient and flexible supervisory method. This effectively compensates for insufficient coverage of fixed monitoring points and the low efficiency of manual inspections, and is gradually being widely applied in the field of highway construction supervision. In construction control zones, traffic cones (hereinafter referred to as "cones") are typically used in conjunction with signs, construction vehicles, and personnel for channelization, isolation, and safety warnings.

[0003] However, existing drone video construction monitoring technologies generally adopt a frame-by-frame detection and real-time reporting processing mode. The core logic is to independently identify and push alarms to the construction elements in each frame. The detection results between frames are isolated from each other and lack a mechanism for cross-frame and cross-target temporal accumulation and information aggregation. Specifically, during the continuous inspection process of drones flying along construction sections, the same construction facility will appear repeatedly in multiple consecutive frames. The existing technology of reporting frame by frame results in a large amount of repetitive information flooding the monitoring platform, which is difficult to effectively align with the session-level time window. In the face of massive amounts of scattered alarm data, the supervisors cannot quickly grasp the overall picture and evolution process of the construction site. At the same time, the drone is affected by airflow, attitude adjustment and other factors during flight, and the onboard video has image jitter and perspective shift. For densely deployed cone-like targets, the detection position of the same cone in different frames will jump laterally. In addition, the cones are densely distributed in strips along the road width, and the cumulative effect of jitter makes the reconstruction of the cone deployment boundary seriously distorted. The existing technology lacks effective noise reduction and position correction methods, and it is difficult to stably restore the true deployment position of each construction facility on the road surface. In addition, the detection results of the existing technology are all expressed in image pixel coordinates, and there is a lack of technical means to map the pixel position to a normalized position with the physical meaning of the road width, which cannot form a construction situation description corresponding to the actual road geometry.

[0004] The existence of the above-mentioned problems makes it difficult to fully utilize the technological advantages of drone video monitoring. The regulatory side cannot obtain panoramic materials that cover the entire construction session cycle and can be compared and recorded, resulting in incomplete record-keeping and verification difficulties, which seriously restricts the level of intelligence and work efficiency of construction supervision. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an intelligent monitoring method for road construction based on UAV video streams, which solves the problems mentioned in the background.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a road construction intelligent monitoring method based on UAV video streams, comprising the following steps:

[0007] Step S1: Detect and track construction-related targets from the current frame of the UAV continuous video stream, perform semantic segmentation on the road to obtain road perception results, select the main road from the road perception results and obtain the geometric constraints of the main road. The geometric constraints of the main road include the parameterized expression of the left curb and the parameterized expression of the right curb. Start the construction session according to the construction triggering conditions.

[0008] Step S2: In the frame where the construction session has started, the detected and tracked targets are sequentially filtered by screen height, road domain inclusion, and confidence. The targets that pass all the filters are taken as valid targets. Based on the center point coordinates of the valid targets and the parameterized expressions of the left and right road edges of the main road, the normalized position of the valid targets in the road width direction is calculated, and the time sequence record is stored with the identity of the target as the primary key.

[0009] Step S3: When the duration of consecutive undetected cones reaches the preset duration, or the number of consecutive undetected cones reaches the preset number of frames, the construction session is determined to end. The time sequence records stored during the session are summarized according to the target identity. The representative position of each target in the road width direction and the representative frame sequence in the time axis direction are calculated. The representative position and the representative frame sequence are used as the horizontal and vertical coordinates, respectively, to construct a two-dimensional semantic graph of road width position and time, and generate a construction facility layout diagram.

[0010] Step S4: Extract the normalized position of the cone target and the representative frame order from the two-dimensional semantic graph, construct the cone target point set, first perform density peak clustering and Gaussian smoothing that keeps the representative frame unchanged on the cone target point set in the road width direction, and then perform secondary density clustering and Gaussian smoothing that keeps the representative frame unchanged on the cone target point set in the two-dimensional space composed of road width position and time.

[0011] Step S5: Generate structured statistical results summarized by category, and perform graphic and text layout on the construction facility layout map. Package the construction facility layout map, structured statistical results and spatiotemporal metadata, and report them to the outside world in one session.

[0012] Furthermore, step S1 includes:

[0013] Simultaneously extract target detection information, target tracking information, and road semantic segmentation information from the input continuous video stream. Use the target detection model to identify construction-related elements and output a two-dimensional detection box. Combine the multi-target tracking algorithm to perform cross-frame temporal correlation on the two-dimensional detection box to extract stable target motion trajectories and target identification. At the same time, use the semantic segmentation model to output a road area mask.

[0014] Select a main road instance from the road area mask, extract the outer contour of the main road and simplify it by polygon approximation, determine two effective road edges on both sides of the road width from the simplified edge set according to the geometric constraints related to the horizontal width of the road, and parametrically express the line segments of the two effective road edges determined by the two endpoints in the image pixel coordinate system to obtain the parametric expressions of the left road edge and the right road edge.

[0015] A construction session is initiated when at least one construction sign is detected, or when the number of cones exceeds a set threshold.

[0016] Furthermore, both the left and right curb parameterization expressions are linear interpolation expressions that represent the horizontal pixel coordinates as a function of the vertical pixel coordinates. The left curb parameterization expression is determined by the coordinates of the two endpoints of the left curb in the image pixel coordinate system, and the right curb parameterization expression is determined by the coordinates of the two endpoints of the right curb in the image pixel coordinate system.

[0017] Furthermore, in step S2, the image height filtering is to remove targets whose vertical coordinates do not fall within the preset image height range;

[0018] The road domain includes filtering to remove targets whose center point does not fall within the road width area defined by the left and right boundaries of the main road;

[0019] Confidence filtering removes targets whose detection confidence is less than or equal to a preset effective threshold.

[0020] Further, in step S2, the calculation of the normalized position includes: substituting the ordinate of the target center point into the parameterized expressions of the left and right road edges respectively to obtain the abscissas of the left and right boundaries of the main road at that ordinate; using the difference between the abscissa of the target center point and the abscissa of the left boundary as the lateral offset within the width of the main road; using the difference between the abscissa of the left boundary and the abscissa of the right boundary as the width of the road; and using the ratio of the lateral offset to the width of the road as the normalized position of the target.

[0021] Furthermore, in step S3, the representative position is the average of the horizontal relative positions in all live frames corresponding to each target identity, which serves as the horizontal drawing reference for the target in the two-dimensional semantic graph.

[0022] The representative frame order is the average of all valid record frame numbers corresponding to each target identity, and is used as the vertical time representative value of the target in the two-dimensional semantic graph.

[0023] Furthermore, in step S3, constructing a two-dimensional semantic graph that combines road width location and time includes:

[0024] The valid targets are sorted according to their last occurrence frame number. The last occurrence frame number of the first valid target after sorting is used as the lower time bound, and the maximum value of the average frame number among all valid targets is used as the upper time bound. The overall frame span is calculated.

[0025] Set the total canvas width, road width drawing area width and margins. Use a truncation function to limit the sum of the frame span and the upper and lower reserved margins within a preset height range to determine the canvas height. Then, calculate the vertical pixel mapping amount for each frame.

[0026] The horizontal representative value of each target is mapped to the horizontal coordinate within the road width drawing area, and the vertical representative value of each target is mapped to the vertical coordinate in the canvas.

[0027] Furthermore, in step S4, performing density peak clustering and Gaussian smoothing while preserving the representative frame invariant on the cone target point set in the road width direction includes:

[0028] Extract the normalized position of all cone targets, perform one-dimensional kernel density estimation within the normalized position interval to obtain the density curve, find local peaks and filter weak peaks, assign each cone target point to the nearest peak center, for cone target points belonging to the same cluster zone, use the representative frame of each cone target point as the anchor point, select the neighboring points whose frame difference does not exceed the first-level time window, calculate the smoothed position of the position according to Gaussian weight, update only the position of the position and do not modify the representative frame.

[0029] Furthermore, in step S4, performing secondary density clustering and Gaussian smoothing while preserving the representative frame invariant on the cone target point set in the two-dimensional space formed by road width location and time includes:

[0030] The representative frames of each cone target point are converted into the ratio of the overall frame span to obtain the normalized time. The normalized position of the road width and the normalized time are combined into two-dimensional feature points.

[0031] Define a scale-constrained Euclidean distance, where the road width direction and time direction are normalized using independent neighborhood scales. When the distance between two points is not greater than a preset neighborhood threshold, an effective cluster is formed.

[0032] The target points of the cones within the effective clusters are subjected to secondary Gaussian smoothing according to the second-level time window. The road width position is updated using the second-level time decay parameter. Free points that are not assigned to any effective clusters are marked as noise points.

[0033] This invention provides a method for intelligent monitoring of road construction based on UAV video streams, which has the following advantages:

[0034] This invention decouples detection triggering from trajectory accumulation, constructing a complete processing link from frame-by-frame perception to session-level aggregation, effectively solving the problems of information redundancy and situational fragmentation caused by frame-by-frame fragmented reporting in the prior art. Specifically, this invention uses target identity identifiers as the primary key to perform session-level accumulation and aggregation of cross-frame temporal observations. At the end of the session, it outputs a global construction evolution unfolding diagram and structured statistical results in a single report, fundamentally avoiding information overload and alignment difficulties caused by fragmented instantaneous alarms. This allows the regulatory side to obtain panoramic traceability materials that are fully aligned with the session time window. Addressing the issue of lateral movement in the detection position of traffic cones caused by drone jitter, this invention calculates the normalized position of the traffic cone target under parameterized main road geometric constraints and applies cascaded one-dimensional density peak clustering and two-dimensional scaled Euclidean neighborhood density clustering. This effectively suppresses lateral spatial noise while strictly maintaining the representative frame, ensuring the stable restoration of the traffic cone target's strip-shaped distribution structure and the authenticity of its temporal semantics. Furthermore, this invention accurately maps the tracking sequence in the two-dimensional image into a two-dimensional semantic map with the normalized position of the traffic cone and the time representative value as the horizontal and vertical axes. This transforms the unstructured video stream into a spatiotemporal situation description with clear physical meaning, significantly enhancing the interpretability of the monitoring results. Through the above methods, this invention provides a set of intelligent solutions for highway construction supervision that leaves a complete record and can be compared and verified, effectively improving the efficiency of analysis and judgment on the supervision side. Attached Figure Description

[0035] Figure 1 This is a flowchart of a road construction intelligent monitoring method based on UAV video stream according to the present invention;

[0036] Figure 2 This is a schematic diagram of a cone being detected in a road construction intelligent monitoring method based on UAV video stream according to the present invention.

[0037] Figure 3 This is a schematic diagram illustrating the detection target as a person in the intelligent monitoring method for road construction based on UAV video stream of the present invention.

[0038] Figure 4 This is a schematic diagram of a vehicle as the detection target in a road construction intelligent monitoring method based on UAV video stream according to the present invention.

[0039] Figure 5 This is a schematic diagram of a road construction intelligent monitoring method based on UAV video stream according to the present invention, in which the target to be detected is a signboard.

[0040] Figure 6This is a schematic diagram of the storage structure of the detected target in the intelligent monitoring method for road construction based on UAV video stream of the present invention.

[0041] Figure 7 This is a schematic diagram of the layout of construction facilities in a road construction intelligent monitoring method based on UAV video stream according to the present invention.

[0042] Figure 8 This is a schematic diagram of the frame-by-frame online processing and session-level post-processing in the intelligent monitoring method for road construction based on UAV video streams of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] Example 1

[0045] Please see Figures 1 to 8 This invention provides a method for intelligent monitoring of road construction based on UAV video streams, employing an architecture that combines frame-by-frame online processing with session-level post-processing, and includes the following steps:

[0046] Step S1: Detect and track construction-related targets from the current frame of the UAV's continuous video stream. Perform semantic segmentation on the road to obtain road perception results. Select the main road from the road perception results and obtain its geometric constraints. The geometric constraints of the main road include the parameterized expressions for the left and right curbs. Initiate a construction session based on construction triggering conditions. Utilize the semantic segmentation model to perform pixel-level road perception on the current frame and output a road area mask as the road perception result. Construction triggering conditions include detecting construction signs or detecting that the number of cones exceeds a set threshold. This step aims to construct road width spatial constraints through a low-level visual perception model and trigger construction data recording based on this. This invention is applicable to inspection scenarios where the UAV flies unidirectionally along the construction road. In scenarios where the UAV travels back and forth or hovers, the system performs forward / reverse correction on the frame sequence based on the flight direction identifier before performing temporal mapping. Step S1 includes:

[0047] Step S11: Simultaneously extract target detection information, target tracking information, and road semantic segmentation information from the input continuous video stream. Use the target detection model to identify construction-related elements and output a two-dimensional detection box. Combine this with a multi-target tracking algorithm to perform cross-frame temporal correlation on the two-dimensional detection boxes to extract stable target motion trajectories and target identification. Simultaneously, use the semantic segmentation model to output a road region mask. In this embodiment, the semantic segmentation model can employ a deep learning network based on an encoder-decoder architecture, such as DeepLabV3+, U-Net, or SegFormer, to perform pixel-level road region segmentation on video frames and output a road region mask under the computational constraints of the UAV embedded computing platform. After the semantic segmentation model outputs the road region mask, the system performs connected component analysis on the mask, removing noise regions with areas smaller than a preset minimum road area threshold, and retaining the top N largest connected components as candidate road instances. Specifically, multimodal visual perception is performed simultaneously on the input continuous video stream. On one hand, a target detection model (such as YOLO series models, YOLOv8, YOLOv10, etc.) is used to identify construction-related elements and output two-dimensional detection boxes. The detected construction-related targets include: cone targets (including normal cones and fallen cones), personnel targets (including construction workers and non-construction workers), vehicle targets (including construction vehicles and other small vehicles), and sign targets (including construction signs, speed limit signs, lane narrowing signs, directional signs, and information boards, etc.). Subsequently, a multi-target tracking algorithm (such as the ByteTrack algorithm, which recovers targets with low scores due to occlusion by associating almost all detection boxes) is used to perform cross-frame temporal association on the above detection boxes to ensure that the same physical target maintains a unique identity in consecutive video frames. This allows for the extraction of stable target motion trajectories. Simultaneously, a semantic segmentation model is used to perform pixel-level processing on the same video frame, outputting masks for all road regions in the image. The information, the mask is mainly used to characterize effective construction areas such as highways and ineffective areas such as urban roads.

[0048] Step S12: Select a main road instance from the road area mask, extract the main road's outer contour and simplify it using polygon approximation. From the simplified edge set, determine two effective road edges on both sides of the road width based on geometric constraints related to the road's lateral width. Parametrically represent the line segments of these two effective road edges in the image pixel coordinate system, defined by their two endpoints, to obtain the parametric expressions for the left and right road edges. The image pixel coordinate system has its origin at the top left corner of the image, with the positive x-axis pointing horizontally to the right and the positive y-axis pointing vertically downwards. Specifically, from the simplified polygon edge set, select the longest... The two longest edges are selected as candidate roadsides. The projected lengths of these two candidate edges in the vertical direction of the image are calculated. If the projected lengths are both greater than a preset ratio (e.g., greater than 80% of the image height), and the average distance between the two candidate edges in the horizontal direction is within a preset road width range, then they are determined as valid roadsides. Otherwise, the remaining edges are traversed in descending order of edge length until the above constraints are met. Specifically, to adapt to the inspection conditions of UAVs flying along construction roads, multiple roads or areas with similar road widths often appear simultaneously in a single frame of top-down image. If the area of ​​interest is not limited, subsequent monitoring of construction elements is easily interfered with by irrelevant road widths. To this end, this invention first obtains several road instances through semantic segmentation. Each instance corresponds to a binary mask and its spatial range within the region of interest, thereby expressing the road region in the form of a calculable set of pixels.

[0049] Under the aforementioned multiple examples, this invention automatically determines the main road instance for subsequent monitoring according to predetermined rules. This ensures that the detection and tracking of construction signs, cones, personnel, and vehicles prioritize the road actually to be monitored in the current frame, reducing interference from adjacent roads or background road widths. In this embodiment, the rightmost road mask in the horizontal direction of the image is selected as the main road to meet the common aerial photography composition requirement that construction-occupied road areas are concentrated on one side of the frame and need to be distinguished from opposite or unconstructed road widths. To further improve inter-frame consistency, the main road selection result can remain stable temporally and can be combined with auxiliary strategies such as temporal filtering, area, or confidence to suppress jumps in road instances between frames.

[0050] In the geometric modeling process, the main road mask is first used to extract the outer contour of the road region, forming a closed boundary. This transforms the discrete pixel-level mask into a polygonal geometric object, providing a foundation for road domain determination and boundary intersection calculation. Since the original contour typically has a large number of vertices, directly participating in subsequent calculations would significantly increase computational and storage overhead. Therefore, this invention simplifies the contour using the Douglas-Peucker polygon approximation algorithm: under a preset approximation tolerance or perimeter ratio constraint, detail vertices that contribute little to the overall shape are merged, allowing the contour to maintain the main road orientation and road width characteristics within an acceptable distortion range, while significantly reducing the number of vertices. This simplified contour is crucial for achieving real-time or near-real-time processing under low computational power conditions and serves as the unified input for subsequent selection of two valid roadsides representing both sides of the road width from the contour edge set, performing road domain inclusion filtering, and road width normalization calculations.

[0051] After simplifying the outline, the present invention proceeds to the roadside determination and parameterized representation stage: from the set of edges of the simplified polygon, based on geometric constraints related to the lateral width of the road, two effective roadside edges located on both sides of the road width are determined; in one embodiment, these two roadside edges extend approximately longitudinally in the image, which can be interpreted as the constraints of the left and right road edges on the road width. To facilitate subsequent unified calculations, the left and right road edges can be described as straight line segments determined by their two endpoints in the image pixel coordinate system, thereby representing the lateral pixel coordinates as a function of the longitudinal pixel coordinates. Let the two endpoints of the left road edge be... and Then the right-side curb can be written as:

[0052]

[0053] Similarly, let the two ends of the right-side curb be... and , can be obtained

[0054]

[0055] Left curb parameterization expression and the parameterized expression for the right curb All are linear interpolation expressions that represent the horizontal pixel coordinates as functions of the vertical pixel coordinates. The parameterized expression for the left curb is derived from the two endpoints of the left curb. , The coordinates of the right curb are determined in the image pixel coordinate system. The parameterized expression for the right curb is derived from the two endpoints of the right curb. , Coordinates are determined in the image pixel coordinate system; , The x and y coordinates of the upper endpoint of the left curb (near the top of the image, i.e., the far end) in the image pixel coordinate system; The x and y coordinates of the lower end of the left curb (near the bottom of the image, i.e., the foreground end) in the image pixel coordinate system; The x and y coordinates of the upper endpoint of the right curb (near the top of the image) in the image pixel coordinate system; The x and y coordinates of the lower end of the right curb (near the bottom of the image) in the image pixel coordinate system; This represents the x-coordinate (pixel coordinate) of the left boundary of the main road at the y-coordinate in the current video frame; y is the y-coordinate in the image pixel coordinate system. This represents the x-coordinate (pixel coordinate) of the right boundary of the main road at the y-coordinate in the current video frame.

[0056] Therefore, the present invention obtains the parameterized representation of the left and right boundaries of the main road within the current video frame. and Based on this, a two-dimensional geometric constraint for the main road is constructed: on the one hand, the construction elements can be filtered for road domain inclusion to determine whether they are located in the effective road surface area; on the other hand, the target horizontal coordinate can be linearly normalized between the left and right boundaries to obtain a normalized position quantity with road width physical meaning, providing a spatial benchmark for subsequent construction element screening and road width-time semantic mapping.

[0057] Step S13: When at least one construction instruction sign is detected, or when the number of cones is greater than a set threshold, start the construction session;

[0058] Session triggering based on main road constraints and complementary dual conditions: During continuous video stream inspection, a primary challenge is the unpredictability of when a construction zone will actually be entered. To avoid blindly recording and processing invalid data from non-construction sections, this invention employs the following triggering logic: focusing only on targets within the geometric constraints of the main road, and establishing two complementary construction triggering conditions:

[0059] Normative requirements (signage): According to the "Construction Management Specifications for Operating Expressways" (DB50 / T 959-2019), construction indicator signs must be laid out in front of the work area as required. Therefore, when this invention detects at least one construction indicator sign, it can serve as a legally mandated standard trigger signal;

[0060] Empirical condition (cone threshold): Based on actual engineering inspection experience, there are often situations where signs are tilted, illegally missing, or obstructed on site, but construction areas will inevitably have a dense deployment of cones. Therefore, when this invention detects that the total number of cones exceeds a set empirical threshold (e.g., more than 3), it can also serve as a fallback trigger signal.

[0061] These two conditions form a logical "OR" relationship, designed to complement each other and greatly enhance the robustness of the invention in complex environments. For example, in actual inspections, if the sign is obscured by passing vehicles or missed due to viewing angle issues, the invention can still be accurately triggered by continuously detected cones; and vice versa. As long as either of the above conditions is met, the invention clearly indicates that "the starting point of the construction area has been reached," and then triggers and enters the "started" session data collection state, officially beginning the recording and accumulation of the tracking sequence.

[0062] Step 2: In frames where the construction session has started, the detected and tracked targets are sequentially filtered through image height, road area inclusion, and confidence levels. Targets that pass all filters are considered valid targets. This includes cone targets, sign targets, personnel targets, and vehicle targets that pass all filters. Based on the center point coordinates of the valid targets and the parametric expressions for the left and right curbs of the main road, the normalized position of the valid targets in the road width direction is calculated. A time-series record is stored using the target's identifier as the primary key. Valid targets are then processed separately according to their category. This step aims to eliminate invalid and erroneous data through a rigorous filtering mechanism and transform the pixel coordinates of high-quality targets into relatively meaningful road surface positions.

[0063] After the construction record is triggered, to prevent the storage of a large amount of invalid or erroneous data, three layers of filtering conditions are set for the tracked target:

[0064] Image height filtering (removal of distortion areas): In order to remove targets whose vertical coordinates do not fall within the preset image height range; since there is often severe perspective distortion at the top and bottom of the image captured by the drone, this invention only retains target data in the middle area where the vertical coordinates occupy 30% to 70% of the total image height.

[0065] Road domain inclusion filtering (removing off-road targets): This is to remove targets whose center point does not fall within the road width area defined by the left and right boundaries of the main road; the center point of the target must strictly fall near the left and right boundaries of the main road determined in step 1, without considering targets on other roads; near the left and right boundaries means that the absolute value of the difference between the x-coordinate of the target center point and the calculated value of the road edge parameterized expression is not greater than the preset road domain tolerance (such as 5 pixels or 2% of the road width).

[0066] Confidence filtering (removing low-quality targets): This removes targets with a detection confidence level less than or equal to a preset effective threshold; the target's detection confidence level must be greater than the preset effective threshold. This condition aims to remove false detections and ensure that the captured image patches have sufficient visual recognition features. In this embodiment, the preset effective threshold can be set to 0.3, where target tracking effectiveness is determined by continuous tracking for at least 3 frames.

[0067] Only high-quality targets that pass the above three layers of filtering will proceed to the next calculation. Targets that pass all three layers of filtering are defined as valid targets, and only valid targets are subjected to subsequent normalized position calculation and storage. In this invention, cone-shaped targets, sign targets, personnel targets, and vehicle targets that pass the above three layers of filtering are collectively referred to as valid targets. This invention calculates the "lateral relative position" of the target on the road based on its distance from the left and right curbs. The specific calculation process is as follows: Let the coordinates of the center point of the current construction element target be... First, set the vertical axis... Substituting the left and right curb fitting parameters determined in step 1, calculate the leftmost boundary x-coordinate of the effective road surface at that height. x-coordinate of the rightmost boundary Then, the lateral relative position of the target is calculated using the following formula:

[0068]

[0069] In the formula, The coordinates of the center point of the current construction element target x-coordinate The leftmost boundary x-coordinate of the effective road surface. The x-coordinate of the rightmost boundary of the effective road surface; in this calculation formula, the y-coordinate of the target center point is substituted into the parameterized expressions of the left and right curbs respectively to obtain the x-coordinates of the left and right boundaries of the main road at this y-coordinate, and the x-coordinate of the target center point is used as the x-coordinate. x-coordinate of the left boundary The difference is used as the lateral offset of the target within the width of the main road, expressed as the right boundary x-coordinate. x-coordinate of the left boundary The difference is used as the road width at the current vertical coordinate. The ratio of the lateral offset to the road width is used as the normalized position of the target. This ratio is between 0 and 1, representing the relative position of the target in the direction of the main road width. The lateral relative position of the target represents the normalized position of the current construction target in the direction of the main road width, with a value of [0,1]. 0 represents the target being located at the leftmost boundary of the road, and 1 represents it being located at the rightmost boundary. After calculating the normalized position, a truncation function is used to restrict it to the closed interval [0,1]. If the normalized position before truncation exceeds the range of [-0.05,1.05], the target is determined to be an outlier outside the road and is removed.

[0070] Finally, the present invention stores the above results. Specifically, the present invention uses a globally unique identifier assigned to the target by the tracking algorithm ( Using the primary key, an independent data record is created for each physical entity appearing in the video. In the corresponding records, this invention stores the following core fields:

[0071] Target Category: Records the category label for this target;

[0072] Temporal and spatial characteristics: Record all frame numbers (FrameNumber) of the target that are validly alive during the tracking period, as well as their corresponding lateral relative positions. Among these steps, recording the frame number is extremely crucial, as this data will directly serve as the basis for calculating the target's vertical coordinate on the spatiotemporal two-dimensional graph in subsequent steps.

[0073] Optimal visual features: To control storage overhead and ensure the readability of subsequent charts, this invention does not save all frames of the target's captured images. Instead, it dynamically compares the images during continuous target tracking, retaining and storing only a cropped image of the frame with the highest detection confidence among all frames, along with its category identifier. This target-based approach... The storage strategy for the aggregation dimension provides data for subsequent spatiotemporal plotting. The specific storage structure is as follows: Figure 6 As shown. After storage, subsequent processing is distributed according to the target category: cone-shaped targets are marked as targets to be smoothed and proceed to step S4 to perform cascaded denoising and smoothing processing; sign targets, personnel targets, and vehicle targets are directly proceeded to step S5 to generate the layout map.

[0074] This invention identifies all selected cone targets, sign targets, personnel targets, and vehicle targets as valid targets and stores their time-series records using the target identifier as the primary key. Only cone targets undergo cascaded denoising and smoothing processing in step S4; sign targets, personnel targets, and vehicle targets directly proceed to layout diagram generation in step S5.

[0075] Step 3: When the duration of consecutive undetected cones reaches a preset duration, or the number of consecutive undetected cone frames reaches a preset number of frames, the construction session is considered to have ended. The time-series records stored during the session are summarized according to the target identity. The representative position of each target in the road width direction and the representative frame sequence in the time axis direction are calculated. Using the representative position and representative frame sequence as the x-axis and y-axis respectively, a two-dimensional semantic graph combining road width position and time is constructed, and a construction facility layout map is generated. This step aims to close the data collection session by automatically identifying the end point of the construction section and transform the accumulated scattered trajectory data into a two-dimensional construction facility layout map that intuitively reflects the actual on-site deployment, serving as the base map for the construction facility layout.

[0076] The specific processing procedure is as follows:

[0077] The construction session termination determination based on detection timeout: When the duration of consecutive undetected cones reaches a preset duration, or the number of consecutive undetected cone frames reaches a preset number of frames, the construction session is determined to end. Specifically, during the UAV inspection process, this invention continuously monitors the detection status of cones. Since the physical facilities in the actual construction area are continuously deployed, in order to accurately delineate the end point of a construction scene, this invention sets a "detection timeout" mechanism: Once no cones are detected for more than the preset duration or preset number of frames, this invention determines that the current construction scene has officially ended, immediately closes the current data collection session, and triggers the subsequent layout diagram generation logic. The preset duration is 10-30 seconds, preferably 15 seconds, and the preset number of frames is 150-450 frames, preferably 200 frames. The starting point of this consecutive undetected period is the frame time of the last successful cone detection.

[0078] Based on the goal Data aggregation and mean calculation: During the plotting preparation phase, iterate through all recorded facility target information within the session to uniquely calculate the mean. Perform feature reduction and dimensionality compression on the unit to obtain robust two-dimensional plotting coordinates:

[0079] Calculate the percentage of construction elements ( (Axis coordinate basics): For each Extract its lateral relative position in all live frames. Calculate its average value As the final horizontal drawing benchmark for this target The horizontal drawing reference is the representative position of the target on the horizontal axis (road width direction) of the two-dimensional semantic map. The mean squared processing can effectively eliminate the lateral position sway caused by drone jitter or single-frame tracking error, ensuring that the lateral placement of the facility on the road width on the map is relatively stable and accurate.

[0080] Calculate the average frame number of construction elements ( (Axis coordinate basics): For each This invention extracts the frame numbers of all valid recorded frames and calculates their average as the time representative value of the target. This time value represents the frame sequence of the target on the vertical axis (time direction) of the two-dimensional semantic graph. Since the drone flies continuously longitudinally along the main road, this frame value can be understood physically as directly mapping to the longitudinal arrangement and relative spacing of the facilities along the road, thus forming the vertical axis reference of the layout diagram. In a video stream with a constant frame rate, the frame sequence is directly mapped to the time axis; in scenarios with frame rate fluctuations or dropped frames, the system converts the frame sequence to physical time based on the timestamps of the video stream, and then uses the physical time as the time axis reference.

[0081] Construction of Construction Facility Layout Diagram with Spatiotemporal Representative Value Extraction and Adaptive Layout: To transform abstract trajectory data into intuitive drawings, a horizontal drawing reference and a time representative value are used as the horizontal and vertical axes, respectively. The spatiotemporal information of each target is mapped onto a two-dimensional canvas. Based on the span of the entity in the frame sequence, an adaptive scaling and image-text separation rendering mechanism is executed. First, the start and end time axes are dynamically defined based on the time representative values ​​of valid targets. This invention arranges all valid targets in ascending order according to their last occurrence frame number, using the last occurrence frame number of the first valid target after sorting as the lower time bound. The timeline starts at the frame number where the first valid target was last observed in the video, and uses the maximum average frame number among all valid targets as the upper bound of the timeline. That is, the maximum average frame number among all targets is used as the endpoint of the time axis. The overall frame span of this mapping is then calculated. The frame span reflects the overall time range of the target's appearance, independent of the physical road chain length. Secondly, adaptive truncation calculation of the canvas height and inverse time mapping rate calculation are performed. This invention presets the total canvas width (denoted as...). ), which includes a fixed-width road width drawing area (denoted as ) and left and right reserved margins When calculating the canvas height, this invention considers the frame span. Plus top and bottom margins Then, strict upper and lower limit thresholds for canvas height are set, denoted as follows: and The final canvas height is dynamically determined using a truncation function. Specifically, the truncation function limits the sum of the frame span and the top and bottom margins to a preset height range to determine the canvas height. The canvas height is: After the canvas height is determined, the vertical pixel mapping amount represented by each frame is calculated backwards, using the following formula: This refers to the number of vertical pixels corresponding to each frame along the time axis; where the top and bottom margins are reserved. The resolution should be 40-100 pixels, preferably 50 pixels; left and right margins should be reserved. The resolution is 40-80 pixels, preferably 60 pixels; The resolution is 400-800 pixels, preferably 500 pixels. The resolution is 2000-4000 pixels, preferably 3000 pixels. Finally, this invention performs coordinate mapping and image-text separation rendering on each facility target. The horizontal representative value of each target is mapped to the horizontal coordinate within the road width drawing area, and the vertical representative value of each target is mapped to the vertical coordinate in the canvas. Specifically, each target... The horizontal representative value (i.e., percentage) ) and longitudinal representative value (i.e., average frame number) Convert to absolute pixel coordinates In terms of lateral spatial mapping, this invention is based on Accurately reconstruct the target's relative position within the road width drawing area, that is, using the target's horizontal representative value. Multiply by the width of the road width drawing area Adding the left margin ΔW, we obtain its horizontal coordinate on the canvas; in terms of vertical time mapping, this invention adopts a drawing logic with time trending upwards, using the formula... Calculate the vertical axis coordinate, that is, the vertical representative value of the target. With the lower bound of time The difference is multiplied by the vertical mapping amount k per frame, and then this value is subtracted from the effective height of the canvas to obtain its vertical coordinate on the canvas. Targets from earlier times are positioned higher vertically on the canvas. During the rendering stage, this invention implements separate anchoring textures, that is, in... Only geometric symbols of the facilities are drawn at the coordinate points. For the optimal confidence capture thumbnail of the target, this invention sets its reference point at a preset offset position on the right side of the road width drawing area, while simultaneously adjusting its ordinate to... Alignment: The thumbnail reference point is set at a fixed offset outside the right boundary of the road width drawing area, with the offset being either 10% of the width of the road width drawing area or a fixed 100 pixels, whichever is greater. During this process, the present invention applies a bounding box-based avoidance mechanism to the thumbnail, automatically shifting it to one side to avoid occlusion, ultimately generating a construction facility layout diagram with clear information display and no interference between the images.

[0082] Step S4: Extract the normalized location and representative frame order of all traffic cone targets from the 2D semantic graph to construct a traffic cone target point set. First, perform density peak clustering and Gaussian smoothing while keeping the representative frames unchanged on the traffic cone target point set in the road width direction. Then, perform secondary density clustering and Gaussian smoothing while keeping the representative frames unchanged on the traffic cone target point set in the 2D space composed of road width location and time. Among them, the density peak clustering in the road width direction adopts the Density Peak Clustering (DPC) algorithm: calculate the local density and relative distance of each traffic cone target point, select the cluster center on the local density-relative distance decision map, and assign non-center points to the cluster to which the point with higher density and closest distance belongs, thereby dividing the clustering zone. To address the issue that traffic cone targets in continuous UAV overhead videos exhibit a strip-like distribution along the road width and that the detection frames are susceptible to lateral movement due to drone shake and changes in viewing angle, this invention performs cascaded denoising and smoothing processing on the traffic cone target point set. Signboard targets, personnel targets, and vehicle targets, due to their sparse numbers or discrete locations, lack clustering statistical significance and are not processed in step S4. Let the representative point of each traffic cone target in the two-dimensional semantic graph be:

[0083]

[0084] in, This indicates the normalized position of the target relative to the width of the main road. This indicates the representative frame order or time value of the target in the session.

[0085] In the road width direction, one-dimensional kernel density estimation is performed on the normalized position of the road width to find peaks, and Gaussian smoothing is performed while keeping the representative frame unchanged. Addressing the issue that cones are distributed in a band-like pattern in the road width direction and are easily affected by detection jitter, this invention divides the road width direction into clusters by density peak clustering, and performs weighted smoothing on the road width position in the temporal neighborhood without changing the temporal representative quantity, thereby suppressing lateral noise and avoiding misrepresenting positional jitter as temporal movement.

[0086] First, extract the normalized location of the road width for all cone targets. A density curve is obtained by performing one-dimensional kernel density estimation within the normalized road width interval. Specifically, evaluation points are generated within the normalized road width interval according to a preset sampling step size, and one-dimensional kernel density estimation is performed to obtain the density curve in the road width direction. The preset sampling step size is 0.005 to 0.02, preferably 0.01. In this embodiment, the one-dimensional kernel density estimation uses a Gaussian kernel function, and the bandwidth is adaptively determined according to the sample standard deviation and sample size using the Silverman empirical rule, or empirically set according to the ratio of the road width normalized interval length to the sample size in a fixed bandwidth mode.

[0087] This invention identifies local peaks on the density curve and filters out weak peaks with relative peak heights below a preset proportion threshold, thereby obtaining one or more cone cluster centers. The preset proportion threshold is 0.3–0.5, preferably 0.4. Subsequently, each cone target point is assigned to the nearest peak center; if the distance between the point and the nearest peak center exceeds a preset assignment distance threshold, it is not forcibly assigned to that cluster to avoid outliers contaminating the smoothing results within the cluster. The preset assignment distance threshold is 0.05–0.15, preferably 0.1.

[0088] For cone-shaped target points that have been grouped into the same cluster zone, this invention uses the representative frame of each cone-shaped target point itself. As anchor points, only neighboring points with frame differences not exceeding the first-level time window are selected for smoothing:

[0089]

[0090] Select neighboring points whose frame difference does not exceed the first-level time window, and calculate the smoothed road width position according to Gaussian weights:

[0091]

[0092] in, The preferred time decay parameter for the first level of smoothing is 8 frames. This represents the time value of the i-th cone target, which is the average of all valid record frame numbers for that target. The time representative value of the j-th neighboring cone target is used to calculate the time difference with the current target. ); The normalized position of the j-th neighboring cone target in the road width direction is used as the object to be weighted averaged; the smoothed road width position is calculated according to Gaussian weights, and only the road width position is updated without modifying the representative frame. This step only updates the road width position of the cone target point. Or its corresponding horizontal pixel coordinates, do not modify Therefore, it will not change the order of the targets on the timeline; among them, the first-level time window is preferably 20 frames.

[0093] In one implementation, the free points that are not assigned to any clustering zone in the one-dimensional stage retain their road width position and are used together with the points after one-dimensional smoothing as the input point set for secondary density clustering; in another implementation, the free points are directly marked as noise points and do not participate in secondary density clustering.

[0094] This invention performs secondary density clustering and Gaussian smoothing while maintaining the representative frame in a two-dimensional space defined by road width location and time. To address the issue that processing only along the road width in one dimension may lead to the erroneous merging of cones that are close in time or space, this invention performs density clustering in a two-dimensional space defined by road width and normalized time, and performs secondary smoothing of road width location within the cluster. This further separates outliers and refines the boundary consistency of the strip structure while maintaining temporal semantics.

[0095] After the first stage of processing, this invention further converts the representative frames of each cone target point into a ratio to the overall frame span, thus obtaining the normalized time:

[0096]

[0097] This allows the normalized location and normalized time of the road width to be combined into two-dimensional feature points:

[0098]

[0099] in, This represents the road width position after the first level of smoothing. To normalize time, we then define a scale-constrained Euclidean distance in this two-dimensional space, where the road width and time directions are normalized using independent neighborhood scales:

[0100]

[0101] in, This represents the neighborhood scale in the road width direction, preferably 0.05. Represents the temporal neighborhood scale, preferably 30 frames. This indicates the session duration. A valid cluster is formed when the distance between two points is no greater than a preset neighborhood threshold. If the number of points is not greater than a preset neighborhood threshold, the two cone-shaped target points are considered to be in the same neighborhood, and density clustering is used to form an effective cluster. The secondary density clustering uses the DBSCAN algorithm: using Euclidean distance with scale constraints as the neighborhood metric, points with no less than the minimum number of clustered points in the neighborhood are marked as core points, and effective clusters are formed from the areas reachable by the density of the core points. The remaining points are marked as noise points. The minimum number of clustered points can be pre-configured according to the application scenario. The preset neighborhood threshold is preferably 1.0, and the minimum number of clustered points is preferably 3. The normalized position of the j-th neighboring cone target in the road width direction is used as the object to be weighted averaged.

[0102] For the target points of the cone that form an effective cluster, the present invention performs secondary Gaussian smoothing again within the cluster according to the second-level time window, and updates the road width position again using the second-level time decay parameter. The second-level time window is preferably 10 frames, and the second-level time window is smaller than the first-level time window. For free points that fail to be assigned to any cluster, they are marked as noise points. These noise points do not participate in the secondary Gaussian smoothing and retain their original drawing position.

[0103] Through the above two-stage processing, the present invention first identifies the strip-shaped clustering structure of traffic cone targets in the road width direction, and then corrects the local discrepancies and misclustering phenomena in the two-dimensional space of road width position-time, thereby suppressing UAV jitter and detection noise.

[0104] Step 5: Generate structured statistical results summarized by category, and perform graphic and text layout on the construction facility layout map. Package the construction facility layout map, structured statistical results, and spatiotemporal metadata, and submit them externally in a single session. This step aims to aggregate all charts and data within the session period in a structured manner to achieve closed-loop reporting for a single inspection task.

[0105] During the monitoring, rendering, and output phase, this invention first performs classification, counting, and state summarization on all valid targets (cone targets, personnel targets, vehicle targets, and sign targets) recorded in the current session, generating corresponding structured statistical results. The state summarization includes: distinguishing between normal and collapsed states for cone targets and counting them separately; distinguishing between construction workers and non-construction workers for personnel targets; distinguishing between construction vehicles and other small vehicles for vehicle targets; and recording the category identifier and detection confidence for sign targets. Subsequently, this invention generates thumbnails of each target of interest in the associated areas of the two-dimensional semantic unfolded diagram to supplement physical details. To further improve the reading experience of the drawings, this invention visualizes the facility density within the scene and enables a text-image avoidance layout mechanism: through overlap detection, iteratively shifting or rearranging conflicting symbol areas and thumbnails to ensure that the final generated raster or vector drawings are clearly distinguishable. The graphic and text avoidance layout mechanism includes: using the smallest bounding rectangle of each symbol area and thumbnail as the bounding box, detecting pixel-level overlap between bounding boxes; if overlap exists, determining the avoidance direction according to a preset priority (thumbnail priority over symbol, lower road target priority over upper road target), and unidirectionally translating outwards from the road drawing area in a fixed step size (e.g., 5 pixels) until there is no overlap or the maximum number of iterations (e.g., 20 times); if overlap still exists after reaching the maximum number of iterations, the conflicting thumbnails are proportionally reduced or hidden. In step S3, the invention retains the cropped small image with the highest detection confidence as a candidate thumbnail for each valid target, and initially anchors its reference point to the offset position on the right side of the layout diagram. In step S5, the invention performs graphic and text avoidance layout on the preliminary layout, iteratively translating, rearranging, or scaling the thumbnails according to the final facility density to generate the final construction facility layout diagram.

[0106] When the video stream ends or the monitoring session triggers the termination condition, this invention formally executes the core closed-loop reporting action. This invention packages the generated construction facility layout map and structured classification statistics table together, and injects spatiotemporal metadata containing information such as session start and end. This spatiotemporal metadata includes: session start frame number, session end frame number, session duration, main road identifier, UAV equipment number, video stream frame rate, and canvas parameters of the two-dimensional semantic graph (total width, road width drawing area, pixel mapping rate). Subsequently, this invention strictly uses a single submission method for output or reporting; and immediately sets an anti-duplicate delivery mechanism upon success, fundamentally avoiding multiple conflicting records from the same construction session to the platform. The anti-duplicate delivery mechanism includes: generating a unique session identifier for each construction session, querying the reporting status bit of this identifier before external reporting; if the status bit indicates reporting has already occurred, skipping this submission; if reporting is successful, setting the status bit and persistently storing it. Through this session-level closed-loop reporting mechanism, the regulatory or operations side no longer receives massive amounts of messy frame-by-frame alarms, but directly obtains objective panoramic materials that are perfectly aligned with the entire session time window. This not only greatly reduces the transmission burden of the link, but also makes subsequent violation tracking, verification and auditing work more intuitive and efficient.

[0107] like Figure 8 As shown, the overall processing flow of this invention is divided into two stages: frame-by-frame online processing and session-level post-processing. In the frame-by-frame online processing stage, the current image is read and decoded frame by frame using the continuous video stream from the UAV as input. Target detection, target tracking, and road semantic segmentation are performed synchronously for each frame to obtain the detection boxes, stable trajectory IDs, and road area mask information of construction-related targets. From the road semantic segmentation results, a main road instance is selected, and the geometric constraint information (parametric expressions for the left and right curbs) of the main road is obtained.

[0108] For frames that have not yet entered a construction session, determine whether to start a construction session based on whether a construction indicator (at least 1) appears in the current frame or whether the number of cones is greater than a set threshold (e.g., ≥3). If any of the above triggering conditions are met, the data collection state is set to started.

[0109] For subsequent frames that have entered the construction session, data acquisition continues, and the detected and tracked targets are sequentially filtered as follows: image region filtering (removing targets whose vertical coordinates do not fall within a preset height range), road domain inclusion filtering (removing targets whose center point does not fall within the road width area defined by the left and right boundaries of the main road), and confidence filtering (removing targets whose detection confidence is not greater than a preset effective threshold). For targets that pass the above three layers of filtering, their normalized position in the road width direction of the main road is calculated, and using the target identity as the primary key, the frame number, category, confidence, road width normalized position, and the thumbnail of the best confidence capture (only the frame with the highest confidence is retained) are appended and saved frame by frame. This frame-by-frame processing process continues to loop until the construction section ends (i.e., the cones are continuously timed out and not detected).

[0110] Next, the session-level post-processing stage begins. First, it is determined whether the current construction session has been triggered and whether there is valid data collected (i.e., whether there is a valid target that has passed the three-layer filtering); if there is no valid data, no report is generated.

[0111] If valid data exists, the time-series records accumulated during the session are summarized according to the target's identity. The representative position of each target in the road width direction (i.e., the average of the lateral relative positions in all surviving frames, used as the horizontal coordinate of the two-dimensional semantic graph) and the representative frame order in the time axis direction (i.e., the average of all valid record frame numbers, used as the vertical coordinate of the two-dimensional semantic graph) are calculated. A two-dimensional semantic canvas is adaptively constructed according to the session time span, and the above representative values ​​are mapped to pixel coordinates in the canvas to form a "road width position-time" two-dimensional semantic graph.

[0112] For targets such as cones that are susceptible to vibration and are distributed in a strip along the road width, a cascaded denoising and smoothing process is further implemented: First, density peak clustering and Gaussian smoothing with representative frames unchanged are performed in the one-dimensional direction of the road width (only the road width position is updated, without modifying the order of representative frames). Then, secondary density clustering and Gaussian smoothing with representative frames unchanged are performed in the two-dimensional space of road width and time (neighborhood determination is performed using scale-constrained scaled Euclidean distance, and secondary Gaussian smoothing is performed on points within the effective clusters) to suppress the impact of UAV vibration and detection errors on the target layout, while strictly ensuring that the time representative quantity is not tampered with.

[0113] After semantic point set optimization, a road base map, target geometric symbols, and guide lines are drawn. Optimal confidence thumbnails of each target are drawn in the right-aligned position of the road width drawing area. The readability of the schematic diagram is improved through a text-image avoidance layout mechanism (iterative translation or rearrangement based on bounding box overlap detection). Simultaneously, structured statistical data is generated by category (classification counts and status summaries of various construction signs, cones, construction vehicles, and personnel).

[0114] Finally, the generated construction evolution diagram, structured statistical data, and spatiotemporal metadata such as session start and end times are packaged and reported externally in a single session, providing construction supervision with a "one-stop" service corresponding to the video time window. Figure 1 The form retains records and a mechanism to prevent duplicate submissions is implemented after successful reporting.

[0115] The specific values ​​of each parameter in this embodiment can be adaptively adjusted according to factors such as the drone's flight altitude, video frame rate, construction scene scale, and target detection model characteristics.

[0116] Some of the data in the above formulas are numerical calculations with dimensions removed, and the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0117] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent monitoring of road construction based on UAV video streams, characterized in that, Includes the following steps: Step S1: Detect and track construction-related targets from the current frame of the UAV continuous video stream, perform semantic segmentation on the road to obtain road perception results, select the main road from the road perception results and obtain the geometric constraints of the main road. The geometric constraints of the main road include the parameterized expression of the left curb and the parameterized expression of the right curb. Start the construction session according to the construction triggering conditions. Step S2: In the frame where the construction session has started, the detected and tracked targets are sequentially filtered by screen height, road domain inclusion, and confidence. The targets that pass all the filters are taken as valid targets. Based on the center point coordinates of the valid targets and the parameterized expressions of the left and right road edges of the main road, the normalized position of the valid targets in the road width direction is calculated, and the time sequence record is stored with the identity of the target as the primary key. Step S3: When the duration of consecutive undetected cones reaches the preset duration, or the number of consecutive undetected cones reaches the preset number of frames, the construction session is determined to end. The time sequence records stored during the session are summarized according to the target identity. The representative position of each target in the road width direction and the representative frame sequence in the time axis direction are calculated. The representative position and the representative frame sequence are used as the horizontal and vertical coordinates, respectively, to construct a two-dimensional semantic graph of road width position and time, and generate a construction facility layout diagram. Step S4: Extract the normalized position of the cone target and the representative frame order from the two-dimensional semantic graph, construct the cone target point set, first perform density peak clustering and Gaussian smoothing that keeps the representative frame unchanged on the cone target point set in the road width direction, and then perform secondary density clustering and Gaussian smoothing that keeps the representative frame unchanged on the cone target point set in the two-dimensional space composed of road width position and time. Step S5: Generate structured statistical results summarized by category, and perform graphic and text layout on the construction facility layout map. Package the construction facility layout map, structured statistical results and spatiotemporal metadata, and report them to the outside world in one session.

2. The intelligent monitoring method for road construction based on UAV video stream according to claim 1, characterized in that: Step S1 includes: Simultaneously extract target detection information, target tracking information, and road semantic segmentation information from the input continuous video stream. Use the target detection model to identify construction-related elements and output a two-dimensional detection box. Combine the multi-target tracking algorithm to perform cross-frame temporal correlation on the two-dimensional detection box to extract stable target motion trajectories and target identification. At the same time, use the semantic segmentation model to output a road area mask. Select a main road instance from the road area mask, extract the outer contour of the main road and simplify it by polygon approximation, determine two effective road edges on both sides of the road width from the simplified edge set according to the geometric constraints related to the horizontal width of the road, and parametrically express the line segments of the two effective road edges determined by the two endpoints in the image pixel coordinate system to obtain the parametric expressions of the left road edge and the right road edge. A construction session is initiated when at least one construction sign is detected, or when the number of cones exceeds a set threshold.

3. The intelligent monitoring method for road construction based on UAV video stream according to claim 2, characterized in that: Both the left and right curb parameterization expressions are linear interpolation expressions that represent the horizontal pixel coordinates as a function of the vertical pixel coordinates. The left curb parameterization expression is determined by the coordinates of the two ends of the left curb in the image pixel coordinate system, and the right curb parameterization expression is determined by the coordinates of the two ends of the right curb in the image pixel coordinate system.

4. The intelligent monitoring method for road construction based on UAV video stream according to claim 1, characterized in that: In step S2, the image height filtering is to remove targets whose vertical coordinates do not fall within the preset image height range; The road domain includes filtering to remove targets whose center point does not fall within the road width area defined by the left and right boundaries of the main road; Confidence filtering removes targets whose detection confidence is less than or equal to a preset effective threshold.

5. The intelligent monitoring method for road construction based on UAV video stream according to claim 4, characterized in that: In step S2, the calculation of the normalized position includes: substituting the ordinate of the target center point into the parameterized expressions of the left and right road edges respectively to obtain the abscissas of the left and right boundaries of the main road at that ordinate; using the difference between the abscissa of the target center point and the abscissa of the left boundary as the lateral offset within the width of the main road; using the difference between the abscissa of the left boundary and the abscissa of the right boundary as the width of the road; and using the ratio of the lateral offset to the width of the road as the normalized position of the target.

6. The intelligent monitoring method for road construction based on UAV video stream according to claim 1, characterized in that: In step S3, the representative position is the average of the horizontal relative positions in all live frames corresponding to each target identity, which serves as the horizontal drawing reference for the target in the two-dimensional semantic graph. The representative frame order is the average of all valid record frame numbers corresponding to each target identity, and is used as the vertical time representative value of the target in the two-dimensional semantic graph.

7. The intelligent monitoring method for road construction based on UAV video stream according to claim 6, characterized in that: In step S3, constructing a two-dimensional semantic graph that combines road width location and time includes: The valid targets are sorted according to their last occurrence frame number. The last occurrence frame number of the first valid target after sorting is used as the lower time bound, and the maximum value of the average frame number among all valid targets is used as the upper time bound. The overall frame span is calculated. Set the total canvas width, road width drawing area width and margins. Use a truncation function to limit the sum of the frame span and the upper and lower reserved margins within a preset height range to determine the canvas height. Then, calculate the vertical pixel mapping amount for each frame. The horizontal representative value of each target is mapped to the horizontal coordinate within the road width drawing area, and the vertical representative value of each target is mapped to the vertical coordinate in the canvas.

8. The intelligent monitoring method for road construction based on UAV video stream according to claim 1, characterized in that: In step S4, performing density peak clustering and Gaussian smoothing while preserving the representative frame on the cone target point set in the road width direction includes: Extract the normalized position of all cone targets, perform one-dimensional kernel density estimation within the normalized position interval to obtain the density curve, find local peaks and filter weak peaks, assign each cone target point to the nearest peak center, for cone target points belonging to the same cluster zone, use the representative frame of each cone target point as the anchor point, select the neighboring points whose frame difference does not exceed the first-level time window, calculate the smoothed position of the position according to Gaussian weight, update only the position of the position and do not modify the representative frame.

9. A method for intelligent monitoring of road construction based on UAV video stream according to claim 8, characterized in that: In step S4, performing secondary density clustering and Gaussian smoothing while preserving the representative frame invariant on the cone target point set in the two-dimensional space consisting of road width location and time includes: The representative frames of each cone target point are converted into the ratio of the overall frame span to obtain the normalized time. The normalized position of the road width and the normalized time are combined into two-dimensional feature points. Define a scale-constrained Euclidean distance, where the road width direction and time direction are normalized using independent neighborhood scales. When the distance between two points is not greater than a preset neighborhood threshold, an effective cluster is formed. The target points of the cones within the effective clusters are subjected to secondary Gaussian smoothing according to the second-level time window. The road width position is updated using the second-level time decay parameter. Free points that are not assigned to any effective clusters are marked as noise points.

Images