Construction site safety monitoring method and system based on AI technology

Abstract

The application belongs to the technical field of safety monitoring, and relates to a construction site safety monitoring method and system based on AI technology, which comprises the following steps: establishing a time-space radar feature baseline map; extracting the global motion trajectory of a specific identification target in an unobstructed visual tracking mode; extracting specific feature data containing the geometric boundary of a corresponding shielding source to generate a context configuration package, and switching to a radar shadow dominant control mode; fusing edge residual echo data to construct corresponding shadow feature descriptions; analyzing the center coordinate set of the cohesive local high-probability connected block to fill in the missing breakpoints of the global motion trajectory; issuing reverse unlocking signaling to return the tracking control right when the double verification true proposition is true; and triggering a gradient collision blocking mechanism. The application solves the problem of track fracture and identity non-continuation of a moving target caused by visual loss under the condition of complex obstacles shielding in a construction site.

Description

Technical Field

[0001] This invention belongs to the technical field of safety monitoring, and relates to a construction site safety monitoring method and system based on AI technology. Background Technology

[0002] In construction sites where heavy machinery interacts with personnel, dynamic safety monitoring is a core element in preventing human-machine collisions. Construction safety monitoring technology faces a common core challenge: how to continuously and uninterruptedly bind the identity and trajectory of moving targets in unstructured construction site environments with complex physical structures, frequent obstructions, and severe interference from light and dust. Especially when workers wearing safety gear or moving construction vehicles travel behind fixed static obstacles such as tower crane foundations, walls, or large stacks of pipes, traditional visual monitoring systems based on single visible light cameras lose their ability to observe the target, creating monitoring gaps. Within these blind spots, the target's identity information is lost, its movement state is unknown, and if its trajectory intersects with heavy equipment operating on the other side, a collision will occur due to a lack of prediction. Constructing a perception architecture that can seamlessly take over tracking tasks when vision fails and accurately restore the target's identifier upon reappearance is a critical technical problem that urgently needs to be solved in the field of construction safety.

[0003] A common industry solution is to deploy multiple fixed-view cameras, using relay coverage to reduce blind spots. This system relies on feature matching algorithms to perform target handover within the overlapping fields of view of the cameras. Another approach is to introduce sensors based on infrared or lidar, such as the image-based smart construction site safety early warning method and system (CN118230261A). Lidar utilizes its immunity to visible light for detection, while lidar can directly acquire 3D point clouds of the environment. Point cloud clustering and morphological matching are used for initial detection and continued tracking of targets behind obstructed areas. Utilizing deep learning-based target detection and tracking models, such as the single-stage target detection algorithm YOLO, combined with the deep learning-based multi-target tracking algorithm DeepSORT, extracts visual features of targets within the visible image and establishes motion models. This has become a common practical method for handling unobstructed situations. At the signal processing level, statistical modeling of radar reflection intensity or using convolutional neural networks to classify targets in radar point clouds are also directions explored in recent years.

[0004] The implementation of multi-camera relay coverage solutions on construction sites faces numerous constraints. The dynamic nature of construction site environments makes it difficult for cameras to obtain fixed, interference-free optimal installation positions. When a target is completely obscured without any overlapping field of view, relying solely on visual feature matching cannot solve the continuity problem of identity across blind spots, and visual features themselves are significantly affected by lighting, dirt, and drastic changes in viewing angle. LiDAR-based tracking solutions also have limitations in such scenarios. LiDAR's electromagnetic waves are in the optical band and cannot penetrate solid obstacles such as cement and steel plates; they can only detect the obstacle itself, not moving targets behind it. Existing technologies, when utilizing radar signals, typically treat radar only as a distance or velocity sensor independent of vision, lacking a systematic mechanism to construct an environmental background baseline, extract weak signal distortion features of the target in shadow areas, and calculate them into reliable location information. Most existing vision and radar fusion modes remain at a shallow data alignment level. When the visual target is lost, the system cannot actively trigger the switching of radar detection mode, and cannot transmit the target motion prior information and occlusion source structure information accumulated in the vision tracking stage to the radar processing unit. As a result, the radar can only re-initiate an indiscriminate full-field search in the blind zone, which results in a heavy computational burden and low target picking accuracy.

[0005] Based on the above problems, the present invention solves the problem of trajectory breakage and identity failure of moving targets due to visual loss caused by complex obstacle occlusion conditions at construction sites. Summary of the Invention

[0006] In a first aspect, the present invention provides a construction site safety monitoring method based on AI technology, comprising the following steps: S1. Acquire and analyze the high-frequency millimeter-wave radar echoes of static obstacles on site, extract three-dimensional and temporal statistical characteristics, and establish a spatiotemporal radar feature baseline map. S2. Simultaneously acquire visible light video stream and point cloud cluster data for target detection and extract the global motion trajectory of a specific target under unobstructed visual tracking mode; S3. Calculate the contact overlap probability of a specific identified target corresponding to the global motion trajectory entering the preset physical visual blind zone. When approaching the spatiotemporal limit threshold point, extract specific representation data containing the geometric boundary of the corresponding occlusion source to generate a context configuration package and switch to the radar shadow-dominated control mode. S4. Based on the context configuration package, delineate the physical space concealed response area, collect real-time radar waves and perform a grid differential comparison operation with the spatiotemporal radar feature baseline map, and fuse the edge residual echo data to construct the corresponding shadow feature description. S5. Input the shadow feature description into the mapping network architecture to perform probability flattening inference and construct a two-dimensional probability occupancy grid map. Analyze the center coordinate set of cohesive local high-probability connected blocks to fill the gaps and breaks in the global motion trajectory. S6. When the target is detected to have escaped the occlusion blind zone, the consistency of the two-dimensional probability occupancy grid map position projection and the unidirectional drop trend law of the shadow quality evaluation factor are used for combined verification. When the double verification is true, a reverse unlocking signal is issued to return the tracking control. S7. Integrate the global motion trajectories within the unrestricted period and the occlusion period to form a smooth identity-bound trajectory record. Substitute it into the four-dimensional spatiotemporal system to deduce the envelope intersection volume and intersection time buffer margin of its intersection with the planned operation exploration trajectory of the heavy engineering equipment on site, and trigger the tiered collision blocking mechanism.

[0007] A further aspect of this invention involves extracting three-dimensional and temporal statistical characteristics to establish a spatiotemporal radar feature baseline map, comprising the following steps: High-frequency millimeter-wave radar echoes observed on-site in a state of no dynamic target intrusion are extracted and clustered to generate target radar reflection units on the surface of static obstacles. Multi-attribute parallel computation is performed on the spatial observation data of the target radar reflection unit to independently derive the Gaussian mixture model distribution data of reflection intensity, the micro-vibration broadening data of the Doppler spectrum, and the spatial phase stability parameters in the time dimension. By merging and assembling Gaussian mixture model distribution data, Doppler spectrum micro-vibration broadening data, and spatial phase stability parameters, a spatiotemporal radar characteristic baseline map with a three-dimensional topological dimensional table combination is constructed.

[0008] A further aspect of this invention involves extracting the global motion trajectory of a specific target under unobstructed visual tracking conditions, comprising the following steps: Under the same clock field of view trigger reference point, the captured visible light video stream and point cloud cluster data are matched, and the point cloud cluster data is called to calibrate and generate the spatial three-dimensional coordinate points of the visible light video target detection box; Based on the video feature confidence index, determine that a specific target is in a stable, unobstructed visual tracking state, and obtain the permission for inter-frame scale continuity in the video. By using a time-series filter to access multi-cycle spatial three-dimensional coordinate points, iterative translation tracking is performed to generate a global motion trajectory.

[0009] A further aspect of the present invention involves extracting specific representational data containing the geometric boundaries of the corresponding occlusion source to generate a context configuration package, including the following steps: Analyze the global motion trajectory to extract and confirm the real-time motion vector that represents the movement trend of a specific target; The size distribution of obstacles in the built-in physical scene in the merged spatial measurement mapping is used to calculate the contact overlap probability of the specific target approaching the occluded isolation surface based on the real-time motion vector. At the instantaneous tangent stage where the contact overlap probability approaches the spatiotemporal limit threshold, the unique identifier of the specific target is merged and encapsulated, and the motion vector and the geometric boundary of the occlusion source are finally established to form a context configuration package, which is then distributed to the lower-level instruction stream.

[0010] A further aspect of this invention involves fusing residual echo data from the edges to construct a corresponding shadow feature description, including the following steps: The high-frequency radar main lobe array is guided to point to the concealed response area in the physical space defined by the context configuration packet, and the radar echo group sequence of the grid feedback azimuth is extracted. The radar echo group sequence and the spatiotemporal radar characteristic baseline map are subjected to a traversal grid subtraction and clearing logic operation to separate and analyze the normalized intensity attenuation profile vector and phase gradient anomaly interval distribution column of the depth dimension. Extract the residual echo at the edge of the boundary between the outer contour of the concealed response area in the physical space, and apply a short-time Fourier transform in the time-frequency domain to capture the edge Doppler leakage frequency. By fusing and splicing normalized intensity attenuation profile vectors, phase gradient anomaly interval distribution columns, and edge Doppler leakage frequency points in the same dimension, a shadow feature description is created.

[0011] A further aspect of this invention involves analyzing the center coordinate set of locally high-probability connected regions to fill in gaps in the global motion trajectory, comprising the following steps: Import shadow feature descriptions into a lightweight regression computation mapping network architecture loaded in the embedded firmware, and output the position offset projection matrix that expands into the possible hidden interval of the target. The global grid centroid dispersion weights of the mapped two-dimensional probabilistic occupancy raster are assigned based on the position offset projection matrix according to probability values. The system signal-to-noise ratio parameter and the distributed information entropy parameter of the current detected differential sequence signal are weighted and fused in parallel to generate a shadow quality evaluation factor characterizing reliable fluctuations; Connectivity analysis parses the centroids of locally high-probability connected blocks in the two-dimensional probabilistic occupancy raster graph to derive the center coordinate set, and injects the center coordinate set into the global motion trajectory time-locked space to maintain a coherent travel route.

[0012] A further aspect of the present invention involves issuing a reverse unlocking signal to reclaim tracking control when the double verification test proves true, comprising the following steps: Accept unknown individuals suddenly identified at the edge of the field of view, calculate and generate the corresponding initial position array of the image capture, and establish a background verification access request across logical domains; The initial position array of the captured image is projected onto the latest version of the two-dimensional probability occupancy grid map to verify the overlapping of the positioning, so as to establish the consistent and matching state of the position projection. By comparing the shadow quality evaluation factors within a preset continuous time sliding window, it was confirmed that they exhibit a unidirectional downward trend law of gradual distortion and weakening. If and only if the position projection consistency match state and the unidirectional falling trend law achieve the simultaneous setting logic, the reverse unlock signal is packaged and released to block the shadow mode and the unique identifier control is hard-switched to the main light film and television world.

[0013] A further aspect of the present invention involves triggering a tiered collision blocking mechanism, comprising the following steps: The system connects and splices together multiple trajectory segments of a specific target within the alternating light and dark visual cycle, creating a smooth interpolation structure that records the identity-binding trajectory across the entire monitoring field of view. Extract and feed the planned operation trajectory of the heavy engineering equipment deployed in the field, map and smooth the identity-bound trajectory record, and merge it into the four-dimensional spacetime system abacus calculation. The three-dimensional geometric intersection that may occur when two moving envelope entities are extrapolated over time is simulated in a three-dimensional manner. The maximum upper limit peak value of the intersection volume of the dynamically connected envelopes is obtained, and the number of seconds until the collision is calculated is used to extract the intersection time buffer margin.

[0014] A further aspect of the present invention, which triggers a tiered collision blocking mechanism, includes the following steps: Based on the decision matrix, the extracted envelope intersection volume is compared and the range of physical destruction extreme values ​​of the collision is determined. Read the intersection time buffer margin assessment of the verification and calculation model, and the tolerance of the last response available for emergency rescue intervention; For field control components that are assessed and confirmed to be approaching the critical level of severe alert, a red-yellow grade forced lock-up braking stop command is issued, ordering the cessation of mechanical activity; Simultaneously, directional loudspeaker warning channels located in the risk intersection area are activated, while pressure is applied to the vibration avoidance warning module of the construction wearable protective gear that penetrates to individual soldiers.

[0015] Secondly, this invention provides a construction site safety monitoring system based on AI technology, comprising the following modules: The spatiotemporal radar feature baseline map construction module is used to acquire and analyze high-frequency millimeter-wave radar echoes of static obstacles on site, extract three-dimensional and temporal statistical characteristics, and establish a spatiotemporal radar feature baseline map. The multimodal vision-driven tracking module is used to simultaneously acquire visible light video streams and point cloud cluster data for fusion target detection and extract the global motion trajectory of a specific target under unobstructed visual tracking mode. The context configuration package generation module is used to calculate the contact overlap probability of a specific identified target corresponding to the global motion trajectory entering the preset physical visual blind zone. When approaching the spatiotemporal limit threshold point, it extracts specific representation data containing the geometric boundary of the corresponding occlusion source to generate a context configuration package and switches to the radar shadow-dominated control mode. The radar shadow feature extraction module is used to delineate the physical space concealed response area based on the context configuration packet, collect real-time radar waves and perform a grid differential comparison operation with the spatiotemporal radar feature baseline map, and fuse edge residual echo data to construct the corresponding shadow feature description. The Shadow Probability Raster Drawing and Trajectory Continuation Module is used to input shadow feature descriptions into the mapping network architecture to perform probability flattening inference and construct a two-dimensional probability occupancy raster map. It also analyzes the center coordinate set of cohesive local high-probability connected blocks to fill in the gaps and breaks in the global motion trajectory. The identity reverse unlock verification module is used to reconstruct the detection of a target escaping the blind spot. It uses the consistency of the two-dimensional probability occupancy grid map position projection and the unidirectional drop trend law of the shadow quality evaluation factor for combined verification. When the double verification is true, it issues a reverse unlock signal to return the tracking control. The continuous safety monitoring and collision blocking triggering module is used to integrate the global motion trajectory within the unrestricted period and the occlusion period to form a smooth identity-bound trajectory record. It is then substituted into the four-dimensional spatiotemporal system to deduce the envelope intersection volume and intersection time buffer margin of the trajectory mapping of the heavy engineering equipment on site, and trigger the tiered collision blocking mechanism.

[0016] In summary, the present invention has the following beneficial technical effects: 1. By constructing a spatiotemporal radar feature baseline map for static obstacles, the Gaussian mixture distribution data, micro-vibration broadening data, and spatial phase stability parameters of each fixed obstacle's reflective unit, obtained during the initial calibration phase under target-free intrusion conditions, are structurally integrated to establish a null environmental background. This mechanism can provide statistically stable baseline feature references for each three-dimensional spatial grid of the environment during subsequent dynamic monitoring. After a target enters the visual blind zone, the system can extract abnormal signals caused by moving targets by performing grid-by-grid differential operations between the real-time radar echo and this pre-stored baseline. This highlights the faint shadow features caused by obstructed human bodies or equipment extracted from the complex static environmental clutter, reducing the probability of false detection and missed detection of dynamic targets in the obstructed environment.

[0017] 2. During the multimodal target perception phase, when the monitoring system determines that the target is in an unobstructed visual tracking state, it implements a resource allocation strategy that actively suppresses the background differential diffraction query thread and limits radar point cloud data to be used only for visual target detection box ranging compensation calibration. This mechanism enables the system to automatically switch to a lightweight radar-assisted mode when there is sufficient visual information, suspending the highly complex radar baseline and real-time signal comparison calculation task, thereby concentrating computing resources on the main path of visual target detection and tracking. It can adaptively adjust the data flow and processing load between multiple sensors according to the current monitoring conditions, avoiding performance redundancy caused by performing full-load sensing analysis in any scenario.

[0018] 3. By analyzing the global motion trajectory to predict the probability of contact overlap, a context configuration package is generated at the instantaneous tangent point stage before the target enters the visual blind zone. This package contains a unique identifier, the final established motion vector, and the complete geometric boundary of the corresponding occlusion source. This package is sent to the underlying radar electrical signal processing system as a state switching exclusive unlock condition. This mechanism transmits the complete spatiotemporal state information and prior knowledge at the moment when visual tracking stops in a structured form from the visual domain to the radar domain. This enables the radar shadow-dominant control unit to focus the radar main lobe onto the defined physical space concealed response area based on the target's terminal motion vector and the exact boundary coordinates of the occlusion source, rather than performing a blind full-field scan. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings are used to provide a further understanding of the present invention.

[0020] Figure 1 A flowchart illustrating an embodiment of this application is disclosed.

[0021] Figure 2 A curve for determining the contact overlap probability as it approaches the spatiotemporal boundary threshold has been published.

[0022] Figure 3 A two-dimensional probability occupancy raster map under occupancy conditions has been disclosed.

[0023] Figure 4 A one-sided trend line chart of the shadow quality evaluation factor has been published.

[0024] Figure 5 A spatial view of the intersection and collision blocking of the continuous safety monitoring envelope volume has been released.

[0025] Figure 6 Structural schematic diagrams of embodiments of this application are disclosed. Detailed Implementation

[0026] The following is in conjunction with the appendix Figure 1 - Figure 6 A preferred description of the present invention is provided below.

[0027] See attached document Figure 1 This invention proposes a construction site safety monitoring method based on AI technology, comprising the following steps: S1. Acquire and analyze the high-frequency millimeter-wave radar echoes of static obstacles on site, extract three-dimensional and temporal statistical characteristics, and establish a spatiotemporal radar feature baseline map. S2. Simultaneously acquire visible light video stream and point cloud cluster data for target detection and extract the global motion trajectory of a specific target under unobstructed visual tracking mode; S3. Calculate the contact overlap probability of a specific identified target corresponding to the global motion trajectory entering the preset physical visual blind zone. When approaching the spatiotemporal limit threshold point, extract specific representation data containing the geometric boundary of the corresponding occlusion source to generate a context configuration package and switch to the radar shadow-dominated control mode. S4. Based on the context configuration package, delineate the physical space concealed response area, collect real-time radar waves and perform a grid differential comparison operation with the spatiotemporal radar feature baseline map, and fuse the edge residual echo data to construct the corresponding shadow feature description. S5. Input the shadow feature description into the mapping network architecture to perform probability flattening inference and construct a two-dimensional probability occupancy grid map. Analyze the center coordinate set of cohesive local high-probability connected blocks to fill the gaps and breaks in the global motion trajectory. S6. When the target is detected to have escaped the occlusion blind zone, the consistency of the two-dimensional probability occupancy grid map position projection and the unidirectional drop trend law of the shadow quality evaluation factor are used for combined verification. When the double verification is true, a reverse unlocking signal is issued to return the tracking control. S7. Integrate the global motion trajectories within the unrestricted period and the occlusion period to form a smooth identity-bound trajectory record. Substitute it into the four-dimensional spatiotemporal system to deduce the envelope intersection volume and intersection time buffer margin of its intersection with the planned operation exploration trajectory of the heavy engineering equipment on site, and trigger the tiered collision blocking mechanism.

[0028] In one embodiment of the present invention, step S1, constructing a spatiotemporal radar feature baseline map for static obstacles to establish a null environmental assumption background, includes the following steps: With no dynamic targets intruding at the construction site, the high-frequency millimeter-wave radar equipment is activated to perform continuous radar beam projection and microwave echo observation of the static obstacle group on site. Target radar reflection units on the surface of static obstacles are extracted from the acquired high-frequency millimeter-wave radar echoes. Gaussian mixture model distribution data of reflection intensity, micro-vibration broadening data of Doppler spectrum, and spatial phase stability parameters are calculated based on the measured attributes of the radar reflection unit using clustering.

[0029] By merging Gaussian mixture model distribution data, Doppler spectrum micro-vibration broadening data, and spatial phase stability parameters, a spatiotemporal radar feature baseline map with a combination of three-dimensional topological features and temporal attribute dimension tables is established. The spatiotemporal radar feature baseline map is stored in the system's resident memory as a reference sample for comparison in abnormal environments.

[0030] Specifically, the system control module switches the safety monitoring system to initialization calibration mode. This mode is triggered when the system starts up or when global monitoring confirms that the construction site is free of dynamic intrusions for a continuous period of time. In this mode, the central processor deployed at a high point on site sends instructions to the high-frequency millimeter-wave radar equipment to initiate baseline environmental scanning of fixed facilities and static obstacles on site. The high-frequency millimeter-wave radar equipment continuously transmits millimeter-wave radar beams at a specific frequency, such as 20Hz, and receives echoes for a duration sufficient to establish a statistically stable background, such as 300s. This process generates a raw echo in-phase and orthogonal dataset (raw echo IQ dataset) containing a large amount of raw echo signals.

[0031] The digital signal processing unit processes the original echo IQ dataset, obtaining a series of range Doppler maps through Fast Fourier Transform (FFT) operations, and extracting high signal-to-noise ratio (SNR) radar point clouds from them. Unsupervised clustering algorithms, such as the density-based noise application spatial clustering algorithm DBSCAN, are used to spatially cluster the radar point clouds, aggregating physically adjacent radar points with similar reflection characteristics into independent static obstacle radar reflection clusters. Each reflection cluster is a target radar reflection unit. For each identified static obstacle radar reflection cluster, the system calculates three-dimensional feature parameters in parallel: First, the radar cross section or signal intensity value of all points in the cluster is extracted within the entire observation time window, and the expectation-maximization algorithm is used to fit these intensity values ​​to obtain Gaussian mixture model distribution data of reflection intensity that can describe its intensity statistical characteristics.

[0032] Second, analyze the distribution of all points in the cluster along the Doppler dimension, calculate the standard deviation or full width at half maximum (FWHM) of the Doppler spectrum, and use this to quantify the minute vibrations caused by environmental factors or the equipment itself (such as engine idling speed), thus forming micro-vibration broadening data of the Doppler spectrum.

[0033] Third, the echo phase variation of each point within the cluster over time is unwrapped, and its phase variance within the observation period is calculated. This phase variance value is defined as the spatial phase stability parameter, and its calculation formula is as follows: ; in, The spatial phase stability parameter is a key indicator for measuring the phase consistency of a reflecting unit. For rigid, static objects, its value approaches 0. Based on experimental data, a stable reference threshold can be set to 0.1 rad. 2 'i' represents the index / number of a specific target radar reflection unit, i.e., an independent "static obstacle radar reflection cluster" obtained through clustering algorithms.

[0034] N represents the number of samples, which is the total number of raw echo data frames acquired by the radar within a set continuous observation period. t represents the time series / frame index, indicating the t-th sample or t-th frame within the observation period. This represents the echo phase value of a specific reflecting unit i at time (or frame) t after unwinding. This represents the average phase of the phase sequence of a specific reflection unit i over the entire observation period (N samples).

[0035] The system structurally integrates the Gaussian mixture model distribution data of reflection intensity, the micro-vibration broadening data of the Doppler spectrum, and the spatial phase stability parameters calculated for each static obstacle radar reflection cluster. This integration process creates a lookup table or hash map indexed by the spatial coordinates or unique identifiers of each reflection cluster, with the corresponding value range containing the aforementioned three sets of characteristic data. This final dataset, possessing a three-dimensional topological structure and containing multi-dimensional signal statistical characteristics, is defined as a spatiotemporal radar characteristic baseline map. This map is then loaded and stored in the system's high-speed random access memory, serving as the baseline null hypothesis background for subsequent steps of environmental change detection and anomalous signal stripping.

[0036] High-frequency millimeter-wave radar equipment refers to millimeter-wave radar operating in the 77GHz to 81GHz frequency band. This band provides centimeter-level ranging accuracy and high angular resolution, making it suitable for fine-grained perception in complex construction sites. The target radar reflection unit is the basic analysis object in this scheme. It does not refer to a single radar reflection point, but rather to a set of point clouds segmented from the original point cloud using clustering algorithms such as DBSCAN, representing a stable reflective surface of a static obstacle (such as a wall or fixed steel structure). The state of no dynamic target intrusion is an operationally defined state, meaning that the system, through a preliminary global scan, has not detected any target moving at a speed exceeding 0.1m / s for 60 consecutive seconds.

[0037] The model order K in the Gaussian mixture model distribution data of reflection intensity is dynamically selected according to the Akaike information criterion, typically ranging from 2 to 4, characterizing the complex reflection intensity distribution caused by multipath effects, etc. Micro-vibration broadening data of the Doppler spectrum is used to capture the inherent vibration characteristics of static objects, with a value range set between 0.01 Hz and 0.2 Hz. This value is based on long-term statistical observations of vibrations generated by typical engineering machinery under idling conditions. The spatiotemporal radar characteristic baseline map is essentially a structured database, with a data structure of {Cluster_ID:[GMM_params, Doppler_std, Phase_variance]}.

[0038] For example, assuming a fixed concrete wall exists, the system observes it in initial calibration mode. A high-frequency millimeter-wave radar device continuously scans at a frequency of 20Hz for 300 seconds, acquiring a total of N, or 6000 frames of data. The digital signal processing unit extracts the radar point cloud corresponding to the wall surface from these 6000 frames of data and clusters them into a static obstacle radar reflection cluster, numbered Cluster_ID_C01. For Cluster_ID_C01, the system begins to calculate features, analyzing the signal strength of its 6000 sample points. It finds that its distribution exhibits a clear bimodal characteristic. The system uses a second-order Gaussian mixture model for fitting, obtaining the Gaussian mixture model distribution data of the reflection intensity as {weight π1=0.8, mean u1=-10dBsm, variance ... Weight π² = 0.2, mean u² = -15 dBsm, variance The standard deviation of the Doppler velocity values ​​of these 6000 sample points was calculated, and the micro-vibration broadening data of the Doppler spectrum was found to be 0.08 Hz, reflecting the weak vibration of the wall caused by the surrounding environment.

[0039] Extract the phase sequence of the cluster center point, untangle it, and calculate its mean value. Substitute into the formula: ; The calculated spatial phase stability parameter is 0.04 rad. 2 The system integrates these calculation results and stores them in the spatiotemporal radar feature baseline map, forming a record: {Cluster_ID_C01: This record serves as a reference for subsequent identification of abnormal disturbances in front of or behind the wall.

[0040] In one embodiment of the present invention, step S2, performing multimodal target perception and initiating a vision-driven tracking mode based on the synchronization data state, includes the following steps: The system utilizes front-end camera hardware and high-frequency millimeter-wave radar equipment to simultaneously acquire visible light video streams and point cloud cluster data within the same field of view. It then calls upon the point cloud cluster data to calibrate and extract the spatial three-dimensional coordinates of the visible light video target detection box. The system determines that the tracked target is in an unobstructed visual tracking state within the line-of-sight chain of the visible light video stream, and generates the target's global motion trajectory using inter-frame feature scale matching and spatial three-dimensional coordinates.

[0041] The system actively suppresses background differential diffraction query threads targeting the spatiotemporal radar feature baseline map to reduce computing power consumption, and forcibly limits the signal direction of the point cloud cluster data to only be used for compensation calibration operations that support the ranging dimension of visible light video target detection boxes.

[0042] Specifically, after completing the construction of the spatiotemporal radar feature baseline map in step S1, the safety monitoring system enters a continuous dynamic monitoring phase. The central processing unit deployed on-site triggers the front-end camera hardware and high-frequency millimeter-wave radar equipment to perform synchronous data acquisition via a synchronization clock signal based on the Network Time Protocol (NAT). The front-end camera hardware captures a 1920x1080 pixel visible light video stream at a rate of 30 frames per second, while the high-frequency millimeter-wave radar equipment simultaneously outputs point cloud cluster data containing reflection points of all dynamic and static objects in the environment. The visible light video stream is fed in real time into a target detection model based on a convolutional neural network, such as YOLOv5. This model analyzes the video stream frame by frame, identifies specific targets in the image, such as workers or vehicles, and outputs their visible light video target detection bounding boxes on the two-dimensional image plane.

[0043] The system calls the stored camera intrinsic and extrinsic parameter matrices to project the two-dimensional coordinate range of each visible light video target detection box into three-dimensional space, forming a view frustum. The system retrieves point cloud cluster data acquired at the same timestamp, filters out all radar points falling within this view frustum, and obtains a depth estimate with strong anti-interference capability by calculating the median of these points in the depth direction. This depth value is combined with the two-dimensional center point of the visible light video target detection box, and after coordinate system transformation, the spatial three-dimensional coordinates of the target in the world coordinate system are generated. The system then enters a state determination loop, analyzing the confidence score of the visible light video target detection box and the number of associated radar points to determine whether the target is in an unobstructed visual tracking mode.

[0044] If the confidence score is higher than the preset threshold and the number of associated points is sufficient, it indicates that the target is clearly visible. The system uses an inter-frame matching algorithm based on deep learning features, such as DeepSORT, to continuously associate the identity markers of the same target in consecutive frames. This algorithm combines spatial three-dimensional coordinate points and uses a Kalman filter to predict and update the target's motion state. The state update equation of the filter is as follows: x K-1It is the state vector at time k-1, F is the state transition matrix, which is preset based on the uniform motion model, w K-1 The process noise vector x is obtained by matrix multiplication and vector addition. K Dimensions and x K-1 To maintain consistency, both are combinations of physical quantities describing the target's motion state. The state vector x in the Kalman filter... K Typically a six-dimensional vector , representing the target's position and velocity in three directions in the world coordinate system, respectively.

[0045] Through continuous iteration, the system generates a time series of continuous three-dimensional spatial coordinate points for each stably tracked target; this is the target's global motion trajectory. In this unobstructed visual tracking mode, the system control module establishes a vision-driven tracking mode. Under this mode, the system proactively executes resource optimization strategies, sending instructions to the operating system kernel to suspend the background differential diffraction query thread responsible for performing differential comparisons between radar signals and spatiotemporal radar feature baseline maps, reducing unnecessary computational consumption. The data flow management module is configured to forcibly limit the signal direction of the point cloud cluster data at this time. Its sole purpose is to provide depth information for visible light video target detection boxes, supporting the compensation and calibration of the ranging dimension of the visible light video target detection boxes, and is no longer passed to deeper signal analysis modules.

[0046] The front-end camera hardware is typically an industrial-grade global shutter CMOS camera, also known as a complementary metal-oxide-semiconductor camera, ensuring image blur-free operation even during high-speed motion. Point cloud cluster data refers to the set of data output by the millimeter-wave radar after a single scan, containing information such as the 3D coordinates, velocity, and reflection intensity of each detection point. Synchronization data status refers to ensuring that the timestamp difference between the exposure midpoint of the video frame and the starting point of the radar scan is controlled within 1ms, guaranteeing the accuracy of data fusion. The visible light video target detection bounding box is a data structure containing the coordinates of the top-left and bottom-right corners of the rectangle, the target category (e.g., "person," "vehicle"), and the confidence score provided by the model.

[0047] The criteria for determining the morphology of unobstructed visual tracking are a target detection confidence score greater than 0.9 and at least three associated valid radar points. Inter-frame feature scale matching is a tracking technique that extracts the target's depth features from the image and finds the most similar feature regions in subsequent frames to achieve target association. The global motion trajectory is a data structure assigned a unique target identifier, internally storing a series of 3D coordinate points sorted by timestamp. The background differential diffraction query thread is a high-level processing procedure reserved for subsequent steps; suppressing its operation in vision-dominated mode is a proactive energy-saving strategy.

[0048] For example, in step S1, a spatiotemporal radar feature baseline map containing the concrete wall Cluster_ID_C01 has been constructed. At t=10.0s, a worker wearing a safety helmet enters the camera's field of view, and the system assigns him a unique identifier, Worker_007. At t=10.1s, the front-end camera hardware captures a frame of image, and the target detection model outputs the visible light video target detection box of Worker_007 with a confidence score of 0.95. The system projects this box into three-dimensional space and filters out 5 radar points located within the corresponding view frustum from the synchronized point cloud cluster data. By calculating the median depth of these 5 points, the system combines the two-dimensional image position of Worker_007 with the radar ranging result to calculate its current three-dimensional spatial coordinates as (15.0, 5.0, 1.2) meters. Since the confidence score of 0.95 is higher than the threshold of 0.9, and the number of associated radar points is 5 (greater than 3), it is determined that Worker_007 is in an unobstructed visual tracking mode. The Kalman filter specific to Worker_007 is initialized, and its initial state vector x 10.1 The position component is set to (15.0, 5.0, 1.2), and the velocity component is initialized to (0, 0, 0).

[0049] In the next frame at t=10.133s, the new position of Worker_007 is measured as (15.05, 5.0, 1.2) meters. The Kalman filter uses this observation to update its state, and its state vector x 10.133 The velocity components are updated to reflect the worker's movement along the positive x-axis. This process continues, and the generated coordinate sequences (15.0, 5.0, 1.2), (15.05, 5.0, 1.2), ... are recorded as the global motion trajectory of Worker_007. Given that the system has entered vision-driven tracking mode, the resource management module will reduce the priority of the background differential diffraction query thread for comparing the spatiotemporal radar feature baseline map to the lowest level and place it in a suspended state. The radar data stream is strictly limited to range compensation calibration for Worker_007, thereby optimizing the overall computational resource allocation of the system.

[0050] See attached document Figure 2 In one embodiment of the present invention, step S3, predicting logical contact trends, generating positive locking events, and passing context configuration packages to the shadow analysis layer, includes the following steps: The system analyzes the global motion trajectory to obtain the real-time motion vector of a specific target. It then calculates the contact overlap probability between the target and the visual blind spot behind a stationary obstacle by combining the size distribution of obstacles in the built-in physical scene. When the contact overlap probability approaches the set spatiotemporal limit threshold and the instantaneous tangent point before the visual line of sight is blocked, the system extracts and organizes a dataset containing the unique identifier of the target, the motion vector, and the complete geometric boundary of the corresponding occlusion source to form a context configuration package.

[0051] A context configuration package is sent to the underlying radar electronic signal processing system as an exclusive unlocking condition for state switching, cutting off the control of the upper-level visual data architecture over the measurement weight and forcing the tracking mode to enter the radar shadow-dominated control unit in sequence.

[0052] Specifically, after acquiring the global motion trajectory of a specific target, the system continuously analyzes it to predict potential visual occlusion risks. The motion analysis module performs real-time parsing of the global motion trajectory, calculating the real-time motion vector of the specific target by applying the weighted finite difference method to the nearest three-dimensional spatial coordinate points in the trajectory sequence. This vector represents the target's current instantaneous velocity and direction of motion. The system retrieves pre-loaded physical scene obstacle size distribution information describing the three-dimensional geometry of all static obstacles in the scene from its internal storage. This distribution information provides the precise volume, surface contour, and spatial position of each obstacle.

[0053] The system extrapolates the real-time motion vector and performs collision detection calculations against the size distribution of obstacles in the built-in physical scene. Finally, it calculates the contact overlap probability of a specific target moving along the current path and eventually entering the visual blind spot behind a stationary obstacle. This contact overlap probability is a comprehensive evaluation metric, and its value monotonically increases as the spatiotemporal distance between the target and the boundary of the visual blind spot decreases. The system sets a spatiotemporal boundary threshold and continuously compares the calculated contact overlap probability with this threshold. When the contact overlap probability first reaches or exceeds the spatiotemporal boundary threshold, and the vision system determines that the current frame is the last fully visible frame before the target is occluded, the system defines this moment as the instantaneous cut-off point stage. During this instantaneous cut-off point stage, the system performs a data snapshot operation, extracts and organizes key state data, and encapsulates it into a structured dataset. This dataset is defined as a context configuration package, which contains three core fields: First, a unique identifier representing the specific target, inherited from the global motion trajectory; Second, the motion orientation vector finally calculated and established at this instantaneous tangent point stage is the motion state supported by the last reliable visual measurement; Third, the complete geometric boundary data of the static obstacle that serves as the occlusion source, determined through collision detection.

[0054] The system state control module sends the context configuration packet as an atomic message to the underlying radar signal processing system via inter-process communication mechanisms, such as shared memory or message queues. The tracking mode arbitrator within the radar signal processing system identifies the received context configuration packet as an exclusive unlock condition for state switching. Once the validity of the configuration packet is verified, the arbitrator executes the control handover logic, forcibly severing the link between the upper-layer visual data architecture and the measurement weight control, and instructs the control priority of the tracking mode to pass to the radar shadow-dominant control unit, thereby completing a seamless switching of the tracking mode.

[0055] The specific identification target is the tracked object assigned a unique identifier in step S2. The real-time motion vector is a three-dimensional vector. Its components represent the instantaneous velocity of the target along each axis in the world coordinate system, with units of m / s. The built-in physical scene obstacle size distribution refers to the static 3D scene map constructed from a building information model or through an initial 3D laser scan. The contact overlap probability is a value normalized to the interval 0 to 1, and its calculation comprehensively considers the predicted time t for the target to reach the blind zone boundary according to its current vector movement. pred And the angle θ between the predicted path and the blind zone normal, when t pred The smaller the value and the closer θ is to 90 degrees, the closer the probability value is to 1.

[0056] The spatiotemporal boundary threshold is a trigger threshold set for the probability of contact overlap. Empirically, to ensure timely and accurate prediction, this value is typically set between 0.8 and 0.95. If the application scenario requires high sensitivity, it can be set to 0.85. The instantaneous cut-off point stage marks the critical point at which vision-driven tracking is about to fail. The context configuration package is a standardized data structure, such as a JSON object containing specific fields, ensuring accurate communication between different system modules. A unique identifier is crucial for ensuring the continuity of the target's identity across different tracking modes.

[0057] The final determination of the motion vector is the basis for initial trajectory extrapolation after switching to shadow tracking mode. The complete geometric boundary of the corresponding occlusion source defines the spatial range that subsequent radar shadow analysis needs to focus on. The radar shadow control unit is a dedicated software module of the radar electronic signal processing system, responsible for taking over the target tracking task when vision fails.

[0058] For example, continuing with the tracking of Worker_007, its global motion trajectory shows that it is approaching the concrete wall Cluster_ID_C01 located at x=18.0 meters at a real-time motion vector v=[1,0,0]m / s. At time t=12.5s, the position of Worker_007 is (17.5, 5.0, 1.2) meters. Based on this, the system predicts that it will come into contact with the visual blind spot formed by the wall after 0.5s, and the calculated contact overlap probability is 0.9. This value exceeds the preset spatiotemporal boundary threshold of 0.85, and the system determines that this moment is the instantaneous cutting point stage.

[0059] The system generates a context configuration package containing the following content: {"unique_id":"Worker_007","final_vector":[1.0, 0, 0]","occluder_geometry":[VertexData_for_C01]}. VertexData_for_C01 represents the complete 3D vertex coordinate data of the wall Cluster_ID_C01 retrieved from the obstacle size distribution of the built-in physical scene. This context configuration package is serialized and sent to the radar electronic signal processing system.

[0060] The radar shadow-dominated control unit within the system receives and parses the configuration packet, confirms it as a valid state switching command, activates the control unit, and gains exclusive control over the tracking state of Worker_007. The system stops using subsequent visual detection results to update the global motion trajectory of Worker_007, completing the switch from visual-dominated to radar shadow-dominated tracking mode.

[0061] In one embodiment of the present invention, step S4, retrieving the context configuration packet to delineate the physical space hidden response region to extract shadow feature description, includes the following steps: By scanning the unique identifier contained in the context configuration package and finally establishing the motion vector, the high-frequency radar wave main lobe coverage array is focused on the physical space concealed response zone where the specific identified target has infiltrated behind the geometric boundary of the obstruction source.

[0062] Real-time radar wave group differential scanning results are extracted from the concealed response area in physical space. The results are compared with the spatiotemporal radar characteristic baseline map by performing grid traversal subtraction operation to extract the normalized intensity attenuation profile vector containing multiple depth interval distribution dimensions and the distribution column of phase gradient anomaly intervals with drastic phase changes.

[0063] The residual echoes at the edge of the concealed response zone in physical space are extracted, and a short-time Fourier transform is applied to calculate and derive the edge Doppler leakage frequency points caused by the intervention of moving objects. These frequencies are then fused into the normalized intensity attenuation profile vector and the phase gradient anomaly interval distribution column to complete the establishment of a unique shadow feature description.

[0064] Specifically, upon receiving the sent context configuration packet, the radar shadow control unit begins executing step S4. The control unit parses the context configuration packet, extracts the unique identifier representing the target and the final established motion vector describing its movement trend, and reads the complete geometric boundary data of the obstruction source. Based on the complete geometric boundary and the final established motion vector, the control unit uses geometric calculations to delineate the area immediately behind the geometric boundary of the obstruction source in three-dimensional space. This area is defined as the physical space concealment response zone where the specific identified target has infiltrated. The system sends a command to the high-frequency radar beam controller to precisely align the radar's main lobe coverage array with this physical space concealment response zone using electronic scanning, concentrating energy for high refresh rate scanning of key areas. The signal processing pipeline extracts real-time radar echo data from this physical space concealment response zone and organizes it into a real-time radar intensity-phase map with the same grid structure as the spatiotemporal radar characteristic baseline map.

[0065] The system performs a grid traversal subtraction operation, comparing the real-time radar intensity-phase map values ​​with the corresponding baseline values ​​retrieved from the spatiotemporal radar characteristic baseline map cell by cell. Through this operation, the system can extract two key differential features: First, a normalized intensity attenuation profile vector containing multiple depth intervals, generated by dividing the concealed response area in physical space along the radar line of sight into multiple equidistant depth slices and calculating the average signal intensity attenuation relative to the baseline within each slice. Second, a distribution of anomalous phase gradient intervals with drastic phase changes, consisting of grid cell coordinates where the difference between the real-time phase and the baseline phase stability parameter exceeds a preset statistical threshold.

[0066] To capture more refined dynamic information, the system further extracts the residual echo signal at the boundary between the concealed response zone and the surrounding unobstructed area in physical space. Short-time Fourier transform calculations are applied to the time-series echoes of this edge region to analyze its spectral characteristics. Since the movement of the obscured target causes diffraction and scattering of radar waves, this calculation can derive one or more edge Doppler leakage frequencies caused by the intervention of the moving object. The system fuses this edge Doppler leakage frequency data with the aforementioned normalized intensity attenuation profile vector and phase gradient anomaly interval distribution column, for example, by splicing them into a higher-dimensional composite vector, thus establishing a unique shadow feature description that comprehensively describes the physical characteristics of the target under obscuration.

[0067] The physical space concealment response zone is a dynamically calculated three-dimensional volume of interest. Its size and location are determined in real time based on the final established motion vector and the complete geometric boundary of the obstruction source, ensuring the effective utilization of radar energy. The mesh traversal subtraction operation is the core differential operation in this step. By subtracting the real-time observation value from the statistically stable background baseline, the detection sensitivity for weak shadow signals is improved. The normalized intensity attenuation profile vector is a one-dimensional vector with dimension M set to 8 to 16. Each element represents the average attenuation level within the depth range. Normalization eliminates the influence of uncertainties in range and target radar cross section. The phase gradient anomaly interval distribution column is a set of coordinate points. Its screening threshold is typically set to three times the baseline phase standard deviation when the real-time phase jitter exceeds this threshold, effectively suppressing noise interference.

[0068] Residual echoes refer to weak radar signals appearing at the edge of obstacles outside the main shadow region due to diffraction. Short-time Fourier transform (SFT) is a time-frequency analysis tool that, by setting appropriate window functions and overlap rates, can capture not only weak low-frequency signals (typically within ±5Hz) caused by target subwavelength micro-movements (such as respiratory signs or idling vibrations), but also accurately extract broadband edge Doppler leakage frequencies caused by macroscopic off-road or walking movements of the target entity. Specific shadow feature descriptions are high-dimensional data structures, not single physical quantities, but rather encode and fuse information from three different physical dimensions: intensity attenuation, phase anomalies, and edge Doppler, thus providing highly fault-tolerant multi-dimensional input features for subsequent decoding and state estimation.

[0069] For example, the radar shadowing control unit receives the context configuration packet generated for Worker_007. Based on the geometric data of the wall Cluster_ID_C01 and the final established motion vector [1.0, 0, 0] m / s of Worker_007, it calculates a cubic region 2 meters wide, 2 meters high, and 3 meters deep behind the wall as the physical space concealment response zone for this mission. The radar beam is guided to this region, and the system acquires the real-time radar echo of this region and performs a grid traversal subtraction operation with the data about Cluster_ID_C01 in the stored spatiotemporal radar characteristic baseline map. The calculation result shows that there is a signal attenuation region about 1.5 meters behind the wall.

[0070] The system divides this region into eight slices along the depth direction and calculates the normalized intensity attenuation profile vector as [0.05, 0.1, 0.3, 0.8, 0.7, 0.4, 0.1, 0.05]. The peak appears in the fourth slice, indicating that the target is located in the middle of the region in terms of depth. The system detects that within this attenuation region, several grid cells have real-time phase jitter exceeding 0.6 rad, which is three times the baseline standard deviation sqrt(0.04). The coordinates of these points are recorded, forming a distribution of phase gradient anomaly intervals. The system performs short-time Fourier transform analysis on the radar echoes at the edge of the attenuation region and detects the corresponding edge Doppler leakage frequency point with a velocity of 0.9 m / s.

[0071] The system combines the 8-dimensional normalized intensity attenuation profile vector, the coordinate point set in the phase gradient anomaly interval distribution column, and the velocity information of 0.9 m / s to form a custom shadow feature description for Worker_007 at the current time t=12.6s, and prepares to pass it to the next step for decoding.

[0072] See attached document Figure 3 In one embodiment of the present invention, step S5, decoding the shadow feature description to construct a two-dimensional probability occupancy grid map under occlusion and performing trajectory continuation, includes the following steps: The shadow feature description containing multiple signal parameters is loaded into a pre-trained and distributed lightweight regression computation mapping network architecture to obtain the position offset projection matrix; the centroid dispersion weights of the unique identifier target are calculated according to the matrix numerical distribution and assembled into a two-dimensional probability occupancy grid map reflecting its occlusion orientation trend; the signal-to-noise ratio of the feedback echo and the network information entropy evaluation are retrieved simultaneously to derive the shadow quality evaluation factor.

[0073] Extract the center coordinate set of the local high-probability connected blocks obtained by aggregation calculation within the two-dimensional probability occupancy grid map, and assemble the center coordinate set as the only numerical supply channel to fill the gaps and breaks in the global motion trajectory during the system's blindness and occlusion period, so as to maintain the uninterrupted trajectory.

[0074] Specifically, after establishing a unique shadow feature description, the system calls this description to estimate the state of the occluded target. The system uses the unique shadow feature description as an input vector, loading it into a lightweight regression computation mapping network architecture that has been trained offline and deployed in an embedded computing unit. This network architecture receives high-dimensional shadow feature data and performs forward propagation computation through its internal nonlinear mapping layer. Its final output layer directly generates a two-dimensional matrix, which is the position offset projection matrix. The values ​​of this position offset projection matrix are arranged and reshaped into a two-dimensional probability occupancy grid map with a predetermined grid size. The value of each cell in this grid map represents the probability that the centroid of the uniquely identified target exists at that physical space location; these values ​​are the centroid dispersion weights of the target.

[0075] A parallel evaluation thread is activated, which synchronously retrieves the raw echo data related to the concealed response zone in the physical space from this radar detection, calculates the signal-to-noise ratio of its differential signal, and combines it with the network information entropy calculated from the probability distribution of the two-dimensional probability occupancy grid. A comprehensive shadow quality evaluation factor is then derived through a weighted fusion model. The calculation formula is as follows: ; Among them, Q SQF SNR represents the shadow quality evaluation factor, a scalar between 0 and 1. The closer the value is to 1, the clearer the shadow signal and the higher the reliability of the localization result. shadow Signal-to-noise ratio (SNR) is calculated from differential radar signals. max This is a reference maximum signal-to-noise ratio calibrated based on system performance, such as 30dB. To ensure the consistency of the calculation model, when the concealed response region is affected by strong clutter interference, resulting in a decrease in the SNR of a single extraction... shadow When the value is negative (i.e., the signal threshold is lower than the noise floor), the system performs a truncation operation on it in the calculation logic (i.e., reduces the SNR). shadow =0dB), ensuring the ratio is non-negative. α represents the weighting coefficient, typically set to 0.6, indicating a greater emphasis on the physical reliability of the signal. p i N is the probability value of the i-th grid cell. cell It represents the total number of grid cells.

[0076] The image processing module analyzes the two-dimensional probabilistic occupancy raster map, filters all high-probability cells by setting a probability threshold, and applies a connected component analysis algorithm to these cells to identify clustered local high-probability connected regions. The system calculates the probability-weighted centroid of the largest connected region, thus obtaining two-dimensional coordinates. These coordinates are combined with the depth information parsed from the normalized intensity attenuation profile vector to form the three-dimensional center coordinate set of the target at the current moment. The system marks the center coordinate set as the unique and valid estimate of the target's position within the current visual blindness occlusion period and appends it along with the corresponding timestamp to the end of the global motion trajectory data sequence. This serves as the sole numerical supply channel to fill gaps in the trajectory, thereby maintaining the uninterrupted recording of the target trajectory.

[0077] The lightweight regression computational mapping network architecture is an optimized deep learning model, such as a multilayer perceptron with two hidden layers and 32 neurons per layer. The input layer dimension matches the dimension of the specific shadow feature description, the number of neurons in the output layer equals the total number of grid cells, and a Softmax activation function is used to ensure that the sum of the output probabilities is one. The position offset projection matrix, in engineering implementation, is the original data form of the two-dimensional probability occupancy grid map. Its name describes its function of projecting abstract shadow features into specific spatial location probabilities. The two-dimensional probability occupancy grid map is typically set as a 20x20 grid covering the physical spatial hidden response area, with each cell representing a 0.2m x 0.2m physical area. The network information entropy term measures the uncertainty of the probability distribution; a highly concentrated probability distribution (low entropy) produces a higher measure of determinism.

[0078] The threshold for filtering locally high-probability connected blocks is set to a single grid cell with a probability exceeding five times the average probability. The center coordinate set is crucial data for maintaining trajectory continuity, representing the most probable location calculated based on a physical model without direct observation.

[0079] For example, the system loads the unique shadow feature description generated by Worker_007 at t=12.6s into a pre-trained multilayer perceptron network. The network outputs a 400-dimensional vector, which the system reshapes into a 20x20 two-dimensional probability occupancy grid map, where the probability values ​​are highly concentrated in a certain region. The system then calculates the signal-to-noise ratio (SNR) of the differential signal. shadow The value is 18 dB, and the Shannon entropy of the 2D probability-occupied raster is calculated to be 4.1 bits. Assuming SNR... max 30dB, N cell With a value of 400 (maximum entropy approximately 8.64 bits) and a weight α of 0.6, the shadow quality rating factor is calculated as follows: .

[0080] The system analyzes the raster image, identifies a locally high-probability connected region centered at cell (10, 6), and calculates its weighted centroid. Combining this centroid coordinate with the depth value of 1.5 meters derived from the peak value of the normalized intensity attenuation profile vector, the system determines the center coordinates of Worker_007 in the world coordinate system to be (19.5, 5.1, 1.2) meters. The system appends the data points {(t: 12.6s), (pos: (19.5, 5.1, 1.2))} to the global motion trajectory of Worker_007, successfully filling the first data breakpoint caused by visual occlusion and maintaining the continuity of the trajectory.

[0081] See attached document Figure 4 In one embodiment of the present invention, step S6, which involves capturing and reconstructing the leaked target information and performing a related physical law verification step to trigger identity reverse unlocking signaling, includes the following steps: When a specific target finishes its movement within the obscured area and is about to re-enter the visible light monitoring range, the system begins the handover verification and control return process in step S6. When the target re-emerges within the visible light acquisition range after emerging from the obscuration, the visual regression master node responsible for processing the video stream detects a single candidate target whose identity has not yet been confirmed. This visual regression master node calculates the target's initial 3D position, forming an initial position array for image capture, and uses this data as a payload to initiate a background verification access request to the radar shadow-dominated control unit currently holding tracking control. Upon receiving this request, the radar shadow-dominated control unit initiates a dual verification procedure: The first part of the verification is spatial location verification. The three-dimensional coordinates of the initial position array captured by the image are projected onto the two-dimensional plane where the two-dimensional probability occupancy grid map is updated in real time. Then, it is checked whether the projected point falls into the confidence area composed of high probability value cells in the two-dimensional probability occupancy grid map. If it falls into the confidence area, it is determined that the physical projection orientation is consistent.

[0082] The second part of the verification is signal trend verification. The control unit retrieves the time series data of the shadow quality evaluation factor for the target at the most recent time steps from its internal cache, performs trend analysis on this sequence, and confirms whether the fluctuation amplitude of the factor has a clear unilateral downward trend law. This unilateral downward trend law reflects the physical process of the target entity gradually moving away from the obstacle occlusion, resulting in a weakening of the shadow effect.

[0083] The system enters a strict logical decision gate. Only when both the physical projection orientation consistency and the unilateral downward trend law are simultaneously verified as true can subsequent operations be executed in the background. Under these conditions, the radar shadow-dominant control unit issues an identity confirmation reverse unlocking signal containing a unique identifier. This signal is sent to the tracking mode controller at the system's top level, triggering the final handover of control. Upon receiving this signal, the controller instructs to sever the mapping link of the radar shadow-dominant control unit, cease generating a two-dimensional probability occupancy grid map for the target, and seamlessly return the binding control of the unique identifier to the visual master tracking domain system where the visual regression master node resides, completing a high-confidence tracking mode switch.

[0084] The visual regression master node refers to the core software module in the system responsible for running the visual target detection and tracking algorithm. A single candidate target specifically refers to a target that newly appears at the edge of an occlusion, but whose identity ID has not yet been confirmed through continuous tracking. The initial position array captured in the image is the three-dimensional coordinate point of the candidate target in the world coordinate system. The criterion for determining the consistency of physical projection orientation is that the projection point of the candidate target falls within the convex hull of the region in the probability-occupied raster map where the probability value is greater than 50% of the peak probability. The background verification access request is an internal inter-process communication message used to trigger the identity handover verification process.

[0085] The criterion for determining a unilateral downward trend law is that the value of the shadow quality evaluation factor must have decreased at least four times within the past five consecutive sampling periods, with a total decrease exceeding 30%, ensuring the stability and significance of the trend. The identity verification reverse unlocking signal is the key trigger in the entire handover process; its essence is a Boolean signal carrying the confirmed target ID. The two-dimensional probabilistic occupancy raster mapping link refers to the entire data processing flow from extracting specific shadow feature descriptions to generating the raster map. The visual main tracking domain system refers to the target tracking logic and state machine driven by visual sensor data.

[0086] For example, continuing the scenario from step S5, Worker_007 is continuously tracked by shadows from t=12.6s to t=13.4s. At t=13.5s, the visual regression master node detects a newly appearing humanoid target at the edge of the wall Cluster_ID_C01 and calculates its initial image capture position array as (20.5, 5.15, 1.2) meters. The visual master node sends a background verification access request to the radar shadow-dominant control unit. The control unit performs spatial verification, projecting the coordinate point (20.5, 5.15) onto the latest two-dimensional probability occupancy grid map it generated for Worker_007. The center of the probability peak region of this grid map is located at (20.48, 5.12), and the newly appearing target position highly coincides with this region. The physical projection orientation consistency is determined to be true.

[0087] The control unit performs trend verification, retrieving five consecutive shadow quality evaluation factor values ​​from t=13.1s to t=13.5s, which are [0.60, 0.55, 0.48, 0.35, 0.20]. This sequence shows a monotonically decreasing trend, satisfying the criterion of a unilateral downward trend. Since both verification conditions are true simultaneously, the control unit generates a reverse unlocking signal for Worker_007 and sends it to the mode controller. Upon receiving the signal, the controller disconnects the shadow mapping link of Worker_007, formally binding the unique identifier "Worker_007" to this newly emerging visual target. The system's tracking control of Worker_007 successfully returns to the main visual tracking domain.

[0088] See attached document Figure 5 In one embodiment of the present invention, step S7, arranging a continuous safety monitoring chain spanning the entire horizon spatiotemporal flow and initiating a tiered collision blocking mechanism, includes the following steps: After achieving continuous tracking of a specific target under both visually visible and occluded conditions, the system enters the final risk assessment and intervention phase. The trajectory integration module merges and solidifies the scattered target trajectory data, extracting global motion trajectory segments generated by the vision system within the unrestricted visible main field of view of the specific target, and global motion trajectory segments inferred by the radar shadow-dominated control unit within the physical space concealment response zone. By stitching these two types of trajectory segments, which have different sources but the same identification identifier, together according to their timestamps, the system obtains a complete, seamless, and smoothly identified trajectory record. This record comprehensively reflects the target's continuous movement history throughout the entire monitoring period.

[0089] The collision prediction module initiates a forward-looking risk assessment calculation. This module retrieves newly generated smooth identity-bound trajectory records from the system database and simultaneously retrieves pre-entered on-site heavy engineering equipment's planned operational trajectory mappings. These planned trajectory mappings describe the planned working paths and operating ranges of large equipment such as excavators and tower cranes over a future period. The system then integrates these two sets of trajectory data into a unified four-dimensional spatiotemporal prediction system. This system employs high-order motion models, such as constant turning rate and acceleration models, to extrapolate the future trajectories of workers and equipment. It also calculates the intersection of their trajectory envelopes in three-dimensional space and time, determining the intersection volume and intersection time slice of all potential envelope collision intersection space clusters.

[0090] The risk decision and execution module evaluates the above calculation results, comparing the calculated upper limit of the intersection volume with the preset safe volume threshold, and simultaneously assessing the buffer margin of the intersection time slice from the current moment. Based on the combination of these two indicators, the system triggers a multi-level tiered collision blocking mechanism. If the upper limit of the intersection volume exceeds the danger threshold and the buffer margin of the intersection time slice is less than the emergency braking window, the electromechanical control module connected to the execution device and the control end via the industrial control bus will forcibly issue the highest priority red-yellow level forced brake stop command. If the risk level is slightly lower, the directional loudspeakers installed in the field will be activated to issue directional loudspeaker warnings to the collision risk area.

[0091] The system sends commands via wireless communication to the wearable protective gear worn by the target worker, such as a smart safety helmet, triggering its built-in vibration module to provide a high-frequency vibration avoidance alert. This vibration avoidance alert is a tactile feedback mechanism that provides additional safety for workers who may not hear audible warnings in noisy environments.

[0092] The unrestricted visible backbone field of view period refers to the time period covered by the vision-dominated tracking mode in steps S2 and S6. The physical space concealed response zone period refers to the time period covered by the radar shadow-dominated tracking mode in steps S3 to S5. Smooth identity-bound trajectory recording is achieved by applying techniques such as spline interpolation or Kalman smoothing to process the stitching points, ensuring the continuity of velocity and acceleration at the mode switching points and avoiding abrupt changes. The planned operation trajectory mapping of heavy engineering equipment on site can be obtained from the BIM model of the construction plan or from the GPS and inertial measurement unit records of the equipment. The fourth dimension in the four-dimensional spatiotemporal prediction system is time; the system treats each target as a three-dimensional volume that moves and deforms over time.

[0093] The envelope collision intersection cluster refers to the set of all four-dimensional points in spacetime where the predicted trajectories of two targets overlap within a future time window. The intersection volume characterizes the severity of the collision, while the intersection time slice indicates the urgency of the warning. The red-yellow grade forced brake stop command is an emergency stop protocol conforming to industrial safety standards; red indicates locking all moving parts, while yellow allows for deceleration to a safe stop. Directional loudspeaker warnings utilize phased array speaker technology to project warning messages (such as "Attention! Vehicle approaching from the left rear!") to the specific worker's location, avoiding interference with the entire construction site.

[0094] For example, the system integrates the trajectory of Worker_007, stitching together its visual tracking segment from t=10.0s to t=12.5s, its shadow tracking segment from t=12.6s to t=13.4s, and its visual tracking segment from t=13.5s to the present, forming a complete, smooth, identity-bound trajectory record. The system retrieves the planned operating trajectory mapping of the nearby excavator Excavator_03, which shows that it will perform a 90-degree turn in 3 seconds. The four-dimensional spatiotemporal prediction system receives these two trajectories and predicts that at t+2.5s, the walking path of Worker_007 will overlap with the end of the excavator's turning radius. The system calculates that the upper limit of the intersection volume of the two envelopes at that time is 0.5 cubic meters, and the intersection time slice buffer margin is 2.5s. According to the risk rule base, this situation is judged as a high-level risk.

[0095] The risk decision module triggers a tiered collision prevention mechanism: a yellow-level braking command is sent to the PLC of the excavator Excavator_03 via the CAN bus, requiring it to smoothly decelerate and pause its slewing motion; the directional loudspeaker closest to Worker_007 is activated to play a warning voice message; and a high-frequency vibration command is sent to the smart safety helmet worn by Worker_007 via the LoRa network. This triple protection ensures that potential collision risks are effectively blocked.

[0096] See appendix Figure 6 This invention also proposes a construction site safety monitoring system based on AI technology, including the following modules: The spatiotemporal radar feature baseline map construction module is used to acquire and analyze high-frequency millimeter-wave radar echoes of static obstacles on site, extract three-dimensional and temporal statistical characteristics, and establish a spatiotemporal radar feature baseline map. The multimodal vision-driven tracking module is used to simultaneously acquire visible light video streams and point cloud cluster data for fusion target detection and extract the global motion trajectory of a specific target under unobstructed visual tracking mode. The context configuration package generation module is used to calculate the contact overlap probability of a specific identified target corresponding to the global motion trajectory entering the preset physical visual blind zone. When approaching the spatiotemporal limit threshold point, it extracts specific representation data containing the geometric boundary of the corresponding occlusion source to generate a context configuration package and switches to the radar shadow-dominated control mode. The radar shadow feature extraction module is used to delineate the physical space concealed response area based on the context configuration packet, collect real-time radar waves and perform a grid differential comparison operation with the spatiotemporal radar feature baseline map, and fuse edge residual echo data to construct the corresponding shadow feature description. The Shadow Probability Raster Drawing and Trajectory Continuation Module is used to input shadow feature descriptions into the mapping network architecture to perform probability flattening inference and construct a two-dimensional probability occupancy raster map. It also analyzes the center coordinate set of cohesive local high-probability connected blocks to fill in the gaps and breaks in the global motion trajectory. The identity reverse unlock verification module is used to reconstruct the detection of a target escaping the blind spot. It uses the consistency of the two-dimensional probability occupancy grid map position projection and the unidirectional falling trend law of the shadow quality evaluation factor for combined verification. When the double verification is true, it issues a reverse unlock signal to return the tracking control. The continuous safety monitoring and collision blocking triggering module is used to integrate the global motion trajectory within the unrestricted period and the occlusion period to form a smooth identity-bound trajectory record. It is then substituted into the four-dimensional spatiotemporal system to deduce the envelope intersection volume and intersection time buffer margin of the trajectory mapping of the heavy engineering equipment on site, and trigger the tiered collision blocking mechanism.

[0097] Each of the modules can be implemented in whole or in part through software, hardware, or a combination thereof. It supports hardware embedded in or independent of the processor in the computer device, and also supports software stored in the memory of the computer device, so that the processor can call and execute the operations corresponding to each of the above modules.

[0098] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A construction site safety monitoring method based on AI technology, characterized in that, Includes the following steps: S1. Acquire and analyze the high-frequency millimeter-wave radar echoes of static obstacles on site, extract three-dimensional and temporal statistical characteristics, and establish a spatiotemporal radar feature baseline map. S2. Simultaneously acquire visible light video stream and point cloud cluster data for target detection and extract the global motion trajectory of a specific target under unobstructed visual tracking mode; S3. Calculate the contact overlap probability of a specific identified target corresponding to the global motion trajectory entering the preset physical visual blind zone. When approaching the spatiotemporal limit threshold point, extract specific representation data containing the geometric boundary of the corresponding occlusion source to generate a context configuration package and switch to the radar shadow-dominated control mode. S4. Based on the context configuration package, delineate the physical space concealed response area, collect real-time radar waves and perform a grid differential comparison operation with the spatiotemporal radar feature baseline map, and fuse the edge residual echo data to construct the corresponding shadow feature description. S5. Input the shadow feature description into the mapping network architecture to perform probability flattening inference and construct a two-dimensional probability occupancy grid map. Analyze the center coordinate set of cohesive local high-probability connected blocks to fill the gaps and breaks in the global motion trajectory. S6. When the target is detected to have escaped the occlusion blind zone, the consistency of the two-dimensional probability occupancy grid map position projection and the unidirectional drop trend law of the shadow quality evaluation factor are used for combined verification. When the double verification is true, a reverse unlocking signal is issued to return the tracking control. S7. Integrate the global motion trajectories within the unrestricted period and the occlusion period to form a smooth identity-bound trajectory record. Substitute it into the four-dimensional spatiotemporal system to deduce the envelope intersection volume and intersection time buffer margin of its intersection with the planned operation exploration trajectory of the heavy engineering equipment on site, and trigger the tiered collision blocking mechanism.

2. The construction site safety monitoring method based on AI technology according to claim 1, characterized in that, Extracting three-dimensional and temporal statistical characteristics to establish a spatiotemporal radar feature baseline map includes the following steps: High-frequency millimeter-wave radar echoes observed on-site in a state of no dynamic target intrusion are extracted and clustered to generate target radar reflection units on the surface of static obstacles. Multi-attribute parallel computation is performed on the spatial observation data of the target radar reflection unit to independently derive the Gaussian mixture model distribution data of reflection intensity, the micro-vibration broadening data of the Doppler spectrum, and the spatial phase stability parameters in the time dimension. By merging and assembling Gaussian mixture model distribution data, Doppler spectrum micro-vibration broadening data, and spatial phase stability parameters, a spatiotemporal radar characteristic baseline map with a three-dimensional topological dimensional table combination is constructed.

3. The construction site safety monitoring method based on AI technology according to claim 1, characterized in that, Extracting the global motion trajectory of a specific target under unobstructed visual tracking includes the following steps: Under the same clock field of view trigger reference point, the captured visible light video stream and point cloud cluster data are matched, and the point cloud cluster data is called to calibrate and generate the spatial three-dimensional coordinate points of the visible light video target detection box; Based on the video feature confidence index, determine that a specific target is in a stable, unobstructed visual tracking state, and obtain the permission for inter-frame scale continuity in the video. By using a time-series filter to access multi-cycle spatial three-dimensional coordinate points, iterative translation tracking is performed to generate a global motion trajectory.

4. The construction site safety monitoring method based on AI technology according to claim 1, characterized in that, Extracting specific representation data containing the geometric boundaries of the corresponding occlusion source to generate a context configuration package includes the following steps: Analyze the global motion trajectory to extract and confirm the real-time motion vector that represents the movement trend of a specific target; The size distribution of obstacles in the built-in physical scene is merged into the spatial measurement mapping to calculate the contact overlap probability of the specific target and the target approaching the occluded isolation surface based on the real-time motion vector. At the instantaneous tangent stage where the contact overlap probability approaches the spatiotemporal boundary threshold, the unique identifier of the specific target is merged and encapsulated, and the motion vector and the geometric boundary of the occlusion source are finally established to form a context configuration package, which is then distributed to the lower-level instruction stream.

5. The construction site safety monitoring method based on AI technology according to claim 1, characterized in that, The process of fusing residual echo data at the edges to construct a corresponding shadow feature description includes the following steps: The high-frequency radar main lobe array is guided to point to the concealed response area in the physical space defined by the context configuration packet, and the radar echo group sequence of the grid feedback azimuth is extracted. The radar echo group sequence and the spatiotemporal radar characteristic baseline map are subjected to a traversal grid subtraction and clearing logic operation to separate and analyze the normalized intensity attenuation profile vector and phase gradient anomaly interval distribution column of the depth dimension. Extract the residual echo at the edge of the boundary between the outer contour of the concealed response area in the physical space, and apply a short-time Fourier transform in the time-frequency domain to capture the edge Doppler leakage frequency. By fusing and splicing normalized intensity attenuation profile vectors, phase gradient anomaly interval distribution columns, and edge Doppler leakage frequency points in the same dimension, a shadow feature description is created.

6. The construction site safety monitoring method based on AI technology according to claim 1, characterized in that, The process of analyzing the center coordinate set of locally high-probability connected regions within a cohesive framework to fill in gaps in the global motion trajectory includes the following steps: Import shadow feature descriptions into a lightweight regression computation mapping network architecture loaded in the embedded firmware, and output the position offset projection matrix that expands into the possible hidden interval of the target. The global grid centroid dispersion weights of the mapped two-dimensional probabilistic occupancy raster are assigned based on the position offset projection matrix according to probability values. The system signal-to-noise ratio parameter and the distributed information entropy parameter of the current detected differential sequence signal are weighted and fused in parallel to generate a shadow quality evaluation factor characterizing reliable fluctuations; Connectivity analysis parses the centroids of locally high-probability connected blocks in the two-dimensional probabilistic occupancy raster graph to derive the center coordinate set, and injects the center coordinate set into the global motion trajectory time-locked space to maintain a coherent travel route.

7. The construction site safety monitoring method based on AI technology according to claim 1, characterized in that, When the double check proves true, a reverse unlock signal is issued to reclaim tracking control, including the following steps: Accept unknown individuals suddenly identified at the edge of the field of view, calculate and generate the corresponding initial position array of the image capture, and establish a background verification access request across logical domains; The initial position array of the captured image is projected onto the latest version of the two-dimensional probability occupancy grid map to verify the overlapping of the positioning, so as to establish the consistent and matching state of the position projection. By comparing the shadow quality evaluation factors within a preset continuous time sliding window, it was confirmed that they exhibit a unidirectional downward trend law of gradual distortion and weakening. If and only if the position projection consistency match state and the unidirectional falling trend law achieve the simultaneous setting logic, the reverse unlock signal is packaged and released to block the shadow mode and the unique identifier control is hard-switched to the main light film and television world.

8. The construction site safety monitoring method based on AI technology according to claim 1, characterized in that, Triggering the tiered collision blocking mechanism includes the following steps: The system connects and splices together multiple trajectory segments of a specific target within the alternating light and dark visual cycle, creating a smooth interpolation structure that records the identity-binding trajectory across the entire monitoring field of view. Extract and feed the planned operation trajectory of the heavy engineering equipment deployed in the field, map and smooth the identity-bound trajectory record, and merge it into the four-dimensional spacetime system abacus calculation. The three-dimensional geometric intersection of two moving envelope entities is extrapolated over time to determine the maximum upper limit peak value of the intersection volume of dynamically connected envelopes, and the intersection time buffer margin is extracted by calculating the number of seconds until the collision.

9. The construction site safety monitoring method based on AI technology according to claim 8, characterized in that, Triggering the tiered collision blocking mechanism also includes the following steps: Based on the decision matrix, the extracted envelope intersection volume is compared and the range of physical destruction extreme values ​​of the collision is determined. Read the intersection time buffer margin assessment of the verification and calculation model, and the tolerance of the last response available for emergency rescue intervention; For field control components that are assessed and confirmed to be approaching the critical level of severe alert, a red-yellow grade forced lock-up braking stop command is issued, ordering the cessation of mechanical activity; Simultaneously, directional loudspeaker warning channels located in the risk intersection area are activated, while pressure is applied to the vibration avoidance warning module of the construction wearable protective gear that penetrates to individual soldiers.

10. A construction site safety monitoring system based on AI technology, characterized in that, Includes the following modules: The spatiotemporal radar feature baseline map construction module is used to acquire and analyze high-frequency millimeter-wave radar echoes of static obstacles on site, extract three-dimensional and temporal statistical characteristics, and establish a spatiotemporal radar feature baseline map. The multimodal vision-driven tracking module is used to simultaneously acquire visible light video streams and point cloud cluster data for fusion target detection and extract the global motion trajectory of a specific target under unobstructed visual tracking mode. The context configuration package generation module is used to calculate the contact overlap probability of a specific identified target corresponding to the global motion trajectory entering the preset physical visual blind zone. When approaching the spatiotemporal limit threshold point, it extracts specific representation data containing the geometric boundary of the corresponding occlusion source to generate a context configuration package and switches to the radar shadow-dominated control mode. The radar shadow feature extraction module is used to delineate the physical space concealed response area based on the context configuration packet, collect real-time radar waves and perform a grid differential comparison operation with the spatiotemporal radar feature baseline map, and fuse edge residual echo data to construct the corresponding shadow feature description. The Shadow Probability Raster Drawing and Trajectory Continuation Module is used to input shadow feature descriptions into the mapping network architecture to perform probability flattening inference and construct a two-dimensional probability occupancy raster map. It also analyzes the center coordinate set of cohesive local high-probability connected blocks to fill in the gaps and breaks in the global motion trajectory. The identity reverse unlock verification module is used to reconstruct the detection of a target escaping the blind spot. It uses the consistency of the two-dimensional probability occupancy grid map position projection and the unidirectional falling trend law of the shadow quality evaluation factor for combined verification. When the double verification is true, it issues a reverse unlock signal to return the tracking control. The continuous safety monitoring and collision blocking triggering module is used to integrate the global motion trajectory within the unrestricted period and the occlusion period to form a smooth identity-bound trajectory record. It is then substituted into the four-dimensional spatiotemporal system to deduce the envelope intersection volume and intersection time buffer margin of the trajectory mapping of the heavy engineering equipment on site, and trigger the tiered collision blocking mechanism.

Images