Monitoring system for interior decoration construction behavior
By extracting the visibility confidence of key points of the human skeleton in video frame processing at the indoor decoration construction site and separating noise in the low frequency domain, and combining the rigid topological radius to correct the predicted coordinates, the error in action feature recognition caused by mechanical vibration and occlusion is solved, and the identification of violations with high accuracy is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENZHEN RIGHT ANGLE DESIGN ENG CO LTD
- Filing Date
- 2026-04-27
- Publication Date
- 2026-07-21
Smart Images
Figure CN122435533A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of behavior monitoring technology, and more specifically to a monitoring system for interior decoration construction activities. Background Technology
[0002] The currently common technical approach includes using image acquisition equipment to obtain video frame sequences of the area to be monitored and extracting the coordinates of key points of the human skeleton in the video frames; inputting the coordinate sequence into a spatiotemporal graph convolutional network to realize the recognition and classification of specific human action semantics. This approach has high recognition accuracy in environments with uniform lighting and simple backgrounds. Its logical basis lies in the fact that key points of the human body present a smooth motion trajectory in continuous video frames.
[0003] However, in interior decoration construction, human body parts are often invisible due to frequent obstruction by scaffolding, workbenches, and building materials. To address the feature loss caused by obstruction, even with well-designed monitoring hardware, the dynamic complexity of the construction environment limits the monitoring logic in human feature extraction and motion compensation. For example, Chinese invention patent application CN119850908A discloses a monitoring method and device for obstruction of the object being measured. It determines operational behavior by extracting the wrist node and calculating the relative positional change between the wrist and the object. Existing technologies typically employ time-series extrapolation algorithms based on historical motion patterns, predicting the motion trend of invisible points to complete the skeleton topology. Under this stressful situation, the unique mechanical vibration environment of the construction site reveals the inherent limitations of existing methods. Frequent operation of high-frequency vibration tools such as impact drills or cutting machines can cause severe spatial Gaussian jump noise in the coordinates of key points, even those that are visible. If these visible points carrying high-frequency noise are directly used as topological references to constrain or extrapolate the coordinates of invisible nodes, it will lead to non-anatomical secondary distortions in the spatial dimension of the predicted coordinates. Conventional linear improvement paths, such as increasing the filter order or simply increasing the model depth, are difficult to retain the true physiological motion characteristics while filtering out mechanical vibration noise. They are also prone to causing lag in the feature tensor in the time dimension, resulting in topological tearing of action features in the early stages of data flow. This feature discrepancy caused by the propagation of noise references makes it difficult for the action classification module to lock stable semantic information from the distorted feature flow, leading to misjudgment or underreporting of illegal construction behavior.
[0004] Therefore, the technical problem to be solved by this invention is how to remove high-frequency visual noise and reconstruct the motion feature topology that conforms to anatomical rigid constraints in a non-ideal environment accompanied by strong mechanical interference and high-density physical occlusion. Summary of the Invention
[0005] This invention proposes a monitoring system for interior decoration construction activities, comprising: The video stream acquisition unit is used to acquire continuous video frames of the construction area; The spatiotemporal feature processing unit, connected to the video stream acquisition unit, is used to extract the coordinate sequence of key points of the human skeleton in continuous video frames and calculate the visibility confidence of the target key points. The spatiotemporal feature processing unit includes: a reference correction module, used to acquire the coordinate sequence of visible anchor points that have a topological connection with the target key points when the visibility confidence of the target key points is lower than a preset visibility threshold, and to separate the low-frequency displacement components of the visible anchor point coordinate sequence within the historical sampling window, and to generate low-frequency reference feature coordinates by removing high-frequency noise components with frequencies exceeding a preset cutoff frequency; a spatial constraint module, used to extract the median Euclidean distance between the target key points and visible anchor points in historical valid frames to establish a rigid topological radius; and a coordinate repair module, used to calculate the predicted coordinates of the target key points based on kinematic laws, and to calculate the Euclidean distance between the predicted coordinates and the low-frequency reference feature coordinates; when the Euclidean distance is greater than the rigid topological radius, the vector along the direction vector from the low-frequency reference feature coordinates to the predicted coordinates is truncated with the rigid topological radius as the scalar length to generate a corrected spatial feature tensor. The action recognition module, connected to the spatiotemporal feature processing unit, is used to calculate the similarity between the spatial feature tensor and the preset template to identify violations.
[0006] Preferably, the spatiotemporal feature processing unit is used to obtain the centroid spatial acceleration of all visible key points of the human skeleton, and adjust the frame length of the historical sampling window according to the magnitude of the centroid spatial acceleration; when the magnitude of the centroid spatial acceleration increases, the spatiotemporal feature processing unit shortens the frame length to reduce the filtering delay; when the magnitude of the centroid spatial acceleration decreases, the spatiotemporal feature processing unit increases the frame length to filter out vibration noise caused by construction tools.
[0007] Preferably, the spatiotemporal feature processing unit includes a spatiotemporal graph convolutional network, which is used to construct a human topology map using the coordinate sequence of key points of the human skeleton, and to complete the feature flow of occluded key points in situ by using low-frequency reference feature coordinates, so as to eliminate the action semantic discontinuity caused by scaffolding occlusion.
[0008] Preferably, it also includes an early warning push unit, which is connected to the action recognition module and is used to push an early warning signal indicating construction safety risks to the site management terminal when a violation is detected. The early warning signal includes the location of the key violation point, the type of violation action, and the current construction work station number.
[0009] Preferably, the benchmark correction module includes a bypass control branch, which is used to shorten the number of sampling frames in the historical sampling window or stop separating low-frequency displacement components when the displacement intensity of key points of the human skeleton exceeds a preset safety threshold, so as to capture falling or transient avoidance dangerous behaviors.
[0010] Preferably, the action recognition module is used to calculate the cosine similarity between the spatial feature tensor and each standard action feature tensor in the preset template, and when the cosine similarity is greater than the matching threshold, it outputs the preset violation category judgment result to which the current construction behavior belongs.
[0011] Preferably, the spatiotemporal feature processing unit is used to calculate the coordinates of the geometric center point of the bounding box of the construction tool, and configure the adjacency matrix weights of the local graph convolution operation based on the Euclidean distance between the geometric center point coordinates and the key point at the end of the human skeleton, so as to enhance the feature recognition of the interaction area between the construction tool and the human body.
[0012] Preferably, the spatiotemporal feature processing unit is used to transform the visible anchor point coordinate sequence to the frequency domain using a Fourier transform operator, perform low-pass filtering on the frequency domain signal according to a preset cutoff frequency, and use an inverse Fourier transform operator to restore the filtered signal to the time domain signal in order to extract low-frequency reference feature coordinates.
[0013] Preferably, the video stream acquisition unit acquires real-time video data with a frame rate of 25Hz to 60Hz from the image acquisition device deployed at the construction site, and uses a tracking algorithm to locate the coordinates of key points of the human skeleton.
[0014] The beneficial effects of this invention are: 1. In the monitoring of interior decoration construction activities, the spatial topological connection dimension of the human skeleton is used to establish a rigid topological reference radius by extracting the median Euclidean distance between the target node and the adjacent visible anchor point within the historical effective window. When the target node is identified as being invisible, the unconstrained predicted coordinates generated by the kinematic equation are truncated by spherical projection using this reference radius. This forces the coordinate compensation term to converge within the physical boundary that conforms to the limits of human anatomy. From the perspective of physical mechanism, this avoids the generation of distortion feature tensors such as abnormal limb stretching and reverse joint folding under occlusion conditions, and ensures the topological authenticity of the action semantic logic in the continuous video frame sequence.
[0015] 2. To address the fluctuations in visual feature acquisition caused by high-frequency vibration tools at construction sites, this solution uses built-in anchor point frequency domain decoupling control logic to separate low-frequency components from the coordinate signals of visible anchor points within a historical sliding data window. By eliminating high-frequency coordinate jump vectors exceeding a preset cutoff frequency, low-frequency reference feature coordinates representing the true displacement of the body are extracted. These low-frequency coordinates are then used as the physical center of the topological constraint sphere, effectively blocking the cascading transmission of mechanical vibration noise to the derivation link of the obstructing node, thereby solving the problem of secondary feature distortion caused by Gaussian jumps in the reference point coordinates.
[0016] 3. Combining the dynamic adjustment mechanism of global centroid spatial acceleration, the system adaptively scales the frame length of the historical sliding data window based on the instantaneous motion intensity of all visible nodes of the human skeleton. During the stable construction operation phase, it maintains high noise resistance and actively reduces the filtering delay or bypasses the frequency domain separation step when a nonlinear change in limb displacement is detected. This ensures that the system has both robustness in filtering environmental noise and sensitivity in capturing transient dangerous behaviors such as falls and sudden evasions in complex environments, ultimately achieving a synergistic balance between noise resistance smoothness and real-time motion recognition. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a logical block diagram of the skeleton topology reconstruction and behavior recognition of the present invention; Figure 2 This is a diagram illustrating the overall composition and early warning process architecture of the monitoring system of this invention. Detailed Implementation
[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0020] A monitoring system for interior decoration construction activities, comprising: The video stream acquisition unit is used to acquire continuous video frames of the construction area; The spatiotemporal feature processing unit, connected to the video stream acquisition unit, is used to extract the coordinate sequence of key points of the human skeleton in continuous video frames and calculate the visibility confidence of the target key points. The spatiotemporal feature processing unit includes: a reference correction module, used to acquire the coordinate sequence of visible anchor points that have a topological connection with the target key points when the visibility confidence of the target key points is lower than a preset visibility threshold, and to separate the low-frequency displacement components of the visible anchor point coordinate sequence within the historical sampling window, and to generate low-frequency reference feature coordinates by removing high-frequency noise components with frequencies exceeding a preset cutoff frequency; a spatial constraint module, used to extract the median Euclidean distance between the target key points and visible anchor points in historical valid frames to establish a rigid topological radius; and a coordinate repair module, used to calculate the predicted coordinates of the target key points based on kinematic laws, and to calculate the Euclidean distance between the predicted coordinates and the low-frequency reference feature coordinates; when the Euclidean distance is greater than the rigid topological radius, the vector along the direction vector from the low-frequency reference feature coordinates to the predicted coordinates is truncated with the rigid topological radius as the scalar length to generate a corrected spatial feature tensor. The action recognition module, connected to the spatiotemporal feature processing unit, is used to calculate the similarity between the spatial feature tensor and the preset template to identify violations.
[0021] Preferably, the spatiotemporal feature processing unit is used to obtain the centroid spatial acceleration of all visible key points of the human skeleton, and adjust the frame length of the historical sampling window according to the magnitude of the centroid spatial acceleration; when the magnitude of the centroid spatial acceleration increases, the spatiotemporal feature processing unit shortens the frame length to reduce the filtering delay; when the magnitude of the centroid spatial acceleration decreases, the spatiotemporal feature processing unit increases the frame length to filter out vibration noise caused by construction tools.
[0022] Preferably, the spatiotemporal feature processing unit includes a spatiotemporal graph convolutional network, which is used to construct a human topology map using the coordinate sequence of key points of the human skeleton, and to complete the feature flow of occluded key points in situ by using low-frequency reference feature coordinates, so as to eliminate the action semantic discontinuity caused by scaffolding occlusion.
[0023] Preferably, it also includes an early warning push unit, which is connected to the action recognition module and is used to push an early warning signal indicating construction safety risks to the site management terminal when a violation is detected. The early warning signal includes the location of the key violation point, the type of violation action, and the current construction work station number.
[0024] Preferably, the benchmark correction module includes a bypass control branch, which is used to shorten the number of sampling frames in the historical sampling window or stop separating low-frequency displacement components when the displacement intensity of key points of the human skeleton exceeds a preset safety threshold, so as to capture falling or transient avoidance dangerous behaviors.
[0025] Preferably, the action recognition module is used to calculate the cosine similarity between the spatial feature tensor and each standard action feature tensor in the preset template, and when the cosine similarity is greater than the matching threshold, it outputs the preset violation category judgment result to which the current construction behavior belongs.
[0026] Preferably, the spatiotemporal feature processing unit is used to calculate the coordinates of the geometric center point of the bounding box of the construction tool, and configure the adjacency matrix weights of the local graph convolution operation based on the Euclidean distance between the geometric center point coordinates and the key point at the end of the human skeleton, so as to enhance the feature recognition of the interaction area between the construction tool and the human body.
[0027] Preferably, the spatiotemporal feature processing unit is used to transform the visible anchor point coordinate sequence to the frequency domain using a Fourier transform operator, perform low-pass filtering on the frequency domain signal according to a preset cutoff frequency, and use an inverse Fourier transform operator to restore the filtered signal to the time domain signal in order to extract low-frequency reference feature coordinates.
[0028] Preferably, the video stream acquisition unit acquires real-time video data with a frame rate of 25Hz to 60Hz from the image acquisition device deployed at the construction site, and uses a tracking algorithm to locate the coordinates of key points of the human skeleton.
[0029] Example 1: In an interior decoration construction site with high-density mobile scaffolding and frequent operation of high-frequency vibratory impact drills by construction workers, the system faces the challenge of spatial Gaussian jumps in the coordinates of visible key points due to mechanical vibrations, and feature stream interruptions caused by physical barriers blocking some limb joints. The video stream acquisition unit collects a continuous video frame sequence containing the aforementioned interference. When the visibility confidence of the target key point is determined to be lower than a preset visibility threshold, the benchmark correction module acquires a sequence of visible anchor point coordinates that have a topological connection with the target key point. Within the historical sampling window, the benchmark correction module separates the low-frequency displacement components of the visible anchor point coordinate sequence, removes high-frequency noise components with frequencies exceeding a preset cutoff frequency, and generates low-frequency benchmark feature coordinates. Based on the mechanical interference of connecting nodes in a multi-degree-of-freedom rigid body system, the system... The spatial displacement variance caused by vibration determines the physical laws governing the reliability of node constraints. For construction conditions where key target points are simultaneously topologically connected to multiple visible anchor points, a multi-source anchor point fusion procedure is introduced. The input state is the spatial coordinate sequence of multiple visible anchor points within a historical sampling window. The three-dimensional spatial fluctuation variance of each visible anchor point coordinate sequence is calculated. The reciprocal of all calculated three-dimensional spatial fluctuation variances is used to generate initial signal-to-noise weights. Each initial signal-to-noise weight is divided by the sum of all initial signal-to-noise weights to obtain a normalized weight coefficient. The low-frequency displacement component of each visible anchor point is multiplied by the corresponding normalized weight coefficient, and vector addition is performed to generate the coordinates of the composite constraint sphere center. When physical occlusion causes the number of available visible anchor points to drop to 1, the fusion procedure is bypassed, and the coordinates of a single visible anchor point are set as the rigid constraint benchmark.
[0030] The spatial constraint module extracts the median Euclidean distance between target keypoints and visible anchor points in historical valid frames, establishing a rigid topological radius. The coordinate repair module calculates the predicted coordinates of the target keypoints based on kinematic laws and calculates their Euclidean distance to the low-frequency reference feature coordinates. When the Euclidean distance is greater than the rigid topological radius, the coordinate repair module performs vector truncation along the direction vector from the low-frequency reference feature coordinates to the predicted coordinates, using the rigid topological radius as a scalar length, generating a corrected spatial feature tensor. The repair process follows the formula below: ,in, The corrected coordinates of the target key points. For low-frequency reference feature coordinates, The predicted coordinates of the target key points. As a rigid topological radius, in the actual calculation process, the system pre-establishes a spatial projection image model based on the installation height and pitch angle of the image acquisition device relative to the construction ground. Using the pixel distance between key points of the spine in the main body of the human body as a real-time scaling reference factor, the Euclidean distance in continuous video frames is normalized to a unified physical scale space. This process eliminates the near-large and far-small error caused by three-dimensional spatial perspective projection, ensuring that the extracted median Euclidean distance can accurately represent the anatomical rigid physical boundary of the human skeleton, avoiding feature distortion when people are far away from or close to the camera. The action recognition module calculates the similarity between the corrected spatial feature tensor and the preset template, and outputs the violation behavior recognition result. The displacement component provided by the benchmark correction module and the rigid boundary established by the spatial constraint module work together to block the transmission of high-frequency vibration noise to the inference link of occlusion nodes. Without introducing a three-dimensional depth sensor, the continuity of action semantic recognition is maintained. This scheme converges the originally divergent noise interference into a physical space that conforms to anatomical constraints, solving the technical problem of decreased recognition accuracy caused by skeleton topological distortion in complex dynamic backgrounds.
[0031] Example 2: On a test platform simulating indoor decoration construction conditions, a human skeleton motion video sequence was acquired using an image sensor with a sampling rate of 30Hz and a resolution of 1920x1080. Vibration devices were installed at the ends of the subjects' limbs to simulate mechanical vibration interference with frequencies from 20Hz to 50Hz. The parameter determined in this experiment was the preset cutoff frequency. The setting of this parameter depends on the balance between filtering out spatial displacement jumps caused by mechanical tools and preserving the displacement trajectory of the human torso. When the spectral bandwidth of the monitored signal is mainly in the interference zone above 15Hz, in order to make the movement of the human core topology coherent and avoid signal aliasing, the preset cutoff frequency was set to 5Hz, and a historical sampling window containing 15 sampling points was selected as the processing benchmark, which corresponds to a time span of 0.5s at a sampling rate of 30Hz.
[0032] The experiment was divided into an experimental group using this technical solution and a control group without frequency domain separation. In the baseline setting stage, the experimental group's input raw wrist coordinate sequence contained high-frequency spatial jump noise with an average amplitude of 12.5 pixels. After entering the processing stage, the baseline correction module performed frequency domain transformation on the coordinate components within the historical sampling window, extracting displacement components with energy below 2Hz. The displacement residual in the generated low-frequency baseline feature coordinates was reduced from the original 12.5 pixels to 1.8 pixels. In the control group, because the high-frequency components were not removed, the baseline point fluctuated between consecutive frames, causing the predicted coordinate direction vector deviation angle generated by the subsequent coordinate repair module to reach 45.2 degrees. The spatial constraint module retrieves 100 frames of valid historical data to calculate the Euclidean distance between target keypoints and visible anchor points. The median of the Euclidean distance sequence is used to establish a rigid topological radius of 85.6 pixels. When a simulated subject holds a drill with their elbow joint obscured by a barrier, the coordinate repair module calculates the predicted coordinates of the elbow joint based on a constant angular velocity kinematic model. The calculated Euclidean distance between the predicted coordinates and the low-frequency reference feature coordinates is 112.4 pixels. Since this value is greater than the rigid topological radius, the system initiates vector truncation, projecting the predicted coordinates onto a rigid sphere centered on the low-frequency reference feature coordinates. The formula is then used to... Update the coordinate node matrix, where, The corrected coordinates of the target key points. For low-frequency reference feature coordinates, The predicted coordinates of the target key points. The radius is the rigid topological radius.
[0033] To verify the system's adaptability under different interference intensities, vibration amplitude gradients of 5 pixels, 15 pixels, and 25 pixels were set in the experiment. The data showed that when the vibration amplitude was in the range of 5 to 15 pixels, the recognition accuracy of the action recognition module was in the range of 97.5% to 98.2%. When the vibration amplitude increased to more than 25 pixels, the recognition accuracy dropped to 82.6% due to the aliasing of the energy distribution of the high-frequency jump component and the low-frequency displacement component. This gradient verification process proved that the defined parameter range enabled the benchmark correction module and the spatial constraint module to work synergistically. Within this range, the low-frequency stable benchmark provided the sphere center coordinates for the vector truncation, compensating for the cumulative error caused by the single time-series prediction. The experimental results showed that by eliminating environmental noise interference in the frequency domain and applying rigid constraints in the spatial domain, the system maintained the physical properties of the skeleton feature flow without introducing a three-dimensional depth sensor. When facing occlusion and vibration conditions, the recall rate of the action recognition module increased from 62.4% in the control group to 95.8% in the experimental group, achieving the recognition of violations at the indoor decoration construction site.
[0034] Example 3: In an interior decoration construction scenario involving ceiling opening operations, mechanical vibration interference with frequencies between 20Hz and 50Hz is generated, accompanied by scaffolding obstruction. The system faces the situation where the skeletal feature flow undergoes spatial jumps due to vibration and some joint features are missing. The spatiotemporal feature processing unit acquires a continuous video frame sequence containing the above interference and extracts the coordinate sequence of key points of the human skeleton. To determine the visibility of the target key points, the spatiotemporal feature processing unit calculates the ratio of the maximum response intensity in the key point heatmap to the average intensity of the local area to generate a visibility confidence score. The local area is defined as 7x centered on the coordinates of the pixel with the maximum response in the heatmap. A 7-pixel rectangular window, calibrated based on the image sensor's imaging resolution at a standard workstation distance, is designed to fully cover the pixel area of the Gaussian thermal distribution of key human body points. This allows for the extraction of the mean environmental background noise, excluding the peak response of key points. The system acquires a 50-frame reference video sequence under unobstructed conditions and extracts the arithmetic mean of the peak response intensity of the heatmap. 0.3 times this arithmetic mean is set as a preset visibility threshold. When the confidence value of a target key point in the current frame falls below this threshold, a subsequent repair procedure is triggered. This is based on the physical phenomenon of the difference in frequency band distribution between human torso displacement and mechanical vibration, designed to address the issue of flattened... During the stable construction period, a dynamic frame length adjustment procedure is established to address high-frequency mechanical interference. The input state consists of the visible key point coordinate sequence and the corresponding clock beat. The baseline frame length of the sampling window is set to 30 frames. The adjustment logic is constructed based on a negative exponential decay model. The actual running frame length is equal to the baseline frame length multiplied by the negative exponent of the natural constant. The negative exponent is equal to the ratio of the current centroid spatial acceleration magnitude to the preset safety threshold. The calculated actual running frame length is rounded down to generate historical sampling window parameters. Quantitative indicators are generated through objective difference calculations to replace the fixed window setting. This includes a single-frame capture mode as an alternative operation path to supplement the conventional multi-frame filtering framework during abnormal drop conditions. Based on the distribution law of the difference in acceleration between free fall and normal construction operations in human biomechanics, the preset safety threshold is set at half the standard value of gravitational acceleration, i.e., 4.9 m / s², as the judgment benchmark. The preset cutoff frequency of 4.5 Hz is determined based on the power spectral density distribution of human limb waving, walking and other autonomous physiological activities in the indoor decoration construction scenario. This frequency can completely preserve the movement trajectory characteristics of construction personnel. At the same time, since the high-frequency vibration energy of mechanical tools such as impact drills is mainly concentrated above 20 Hz, this setting achieves accurate separation of the signal frequency domain. The preset safety threshold of 4.9 m / s² is referenced to 0.The 5x setting is the initial stage acceleration physical characteristic value for determining the transition of the human body's center of mass from a static state to an accidental fall state. It is used to distinguish between normal position adjustments during construction operations and sudden fall hazards. The spatiotemporal feature processing unit obtains the three-dimensional coordinate set of all visible key points of the human skeleton and calculates the geometric center to generate the center of mass coordinates. The second-order time difference operation is applied to the center of mass coordinate sequence to extract the spatial acceleration modulus of the center of mass. When the spatial acceleration modulus of the center of mass is greater than the preset safety threshold, the bypass control branch is triggered to stop separating the low-frequency displacement components. When the coordinate repair module determines that the target key point is in an occluded state, it calls the effective coordinate components of the two historical frames. and A first-order constant velocity kinematic model is constructed, and the coordinate repair module calculates the unconstrained predicted coordinates of the current frame based on this model. The specific calculation formula is as follows: ,in, The predicted coordinates of the target key points. The corrected coordinates generated for the previous frame. The model generates corrected coordinates for the first two frames and provides a geometric estimate of the damaged topology by capturing the motion components of nodes in the time dimension.
[0035] The baseline correction module performs frequency domain analysis on the visible anchor point coordinate sequence that has a topological connection with the target key point. The system extracts the frequency band where the energy peak generated by the impact drill vibration is located through fast Fourier transform and identifies it as the high-frequency interference zone. To ensure that the generated low-frequency baseline feature coordinates can reflect the displacement trajectory of the body center, the preset cutoff frequency is set to 4.5Hz. The baseline correction module applies Butterworth low-pass filtering to the coordinate sequence within the historical sampling window to remove high-frequency displacement components exceeding the preset cutoff frequency and generate low-frequency baseline feature coordinates. When dealing with the case where the elbow joint is obstructed, the spatial constraint module establishes the rigid topological radius based on the median of the Euclidean distance sequence between the shoulder anchor point and the elbow node recorded in the historical valid frames. When the Euclidean distance between the predicted coordinates calculated by the coordinate repair module and the low-frequency baseline feature coordinates is greater than the rigid topological radius, the coordinate repair module performs vector truncation along the direction vector from the low-frequency baseline feature coordinates to the predicted coordinates, generating a corrected spatial feature tensor. The truncation process follows the following formula: ,in, The corrected coordinates of the target key points. For low-frequency reference feature coordinates, The predicted coordinates of the target key points. Using a rigid topological radius, the action recognition module calculates the similarity between the corrected spatial feature tensor and the preset behavior template to identify violations such as abnormal drilling height or improper machine holding posture. The stable displacement component generated by the benchmark correction module and the anatomical rigid boundary established by the spatial constraint module work together to block the transmission of high-frequency vibration noise to the occlusion node inference link. Without adding three-dimensional depth sensor hardware, this technical solution maintains the violation behavior recognition recall rate above 94.0% through frequency domain reconstruction and spatial projection constraints of the damaged coordinate sequence.
[0036] After receiving the corrected spatial feature tensor, the action recognition module extracts the feature vector sequence representing the skeleton motion trajectory from the tensor. It also calls the standard action matrix stored in the preset illegal action library. Similarity comparison is performed based on the Gaussian decay law of spatial interaction strength with distance. The spatiotemporal feature processing unit extracts the intersection of the diagonals of the bounding box coordinates of the construction tool to generate the coordinates of the geometric center point. The spatial Euclidean distance between the coordinates of the geometric center point and the coordinates of the key point at the end of the human skeleton is calculated using a distance metric algorithm. The system executes the local graph convolution operation and the adjacency matrix weight quantization allocation procedure. The input state is the predetermined decay parameter and the real-time spatial coordinate system. The weight value is calculated according to the Gaussian kernel function. The weight value is equal to the negative exponent of the natural constant. The corresponding negative exponent value is equal to the square of the spatial Euclidean distance divided by 2 times the preset decay variance. The preset decay variance is obtained through engineering calibration experiments. The image acquisition device extracts the spatial Euclidean distance of 10 frames under static holding operation posture and calculates the statistical variance. This is multiplied by 1.5 times the engineering margin constant to generate the preset decay variance. The weight values are calculated to update the edge connection strength of the spatiotemporal graph convolutional network. When the tool is far away from the human body and the weight value is lower than 0.01, the mean pooling procedure is used to replace the local graph convolution operation to provide alternative logic in extreme edge states. The action recognition module calculates the feature vector sequence. With standard action matrix Cosine similarity between The action matching score is obtained, and the calculation process follows the formula. ,in, The cosine similarity score is used to calculate the similarity score. For real-time extracted motion feature vectors, For the standard violation action matrix, when the similarity score Greater than the preset similarity threshold At that time, the system outputs the corresponding category of violation identification results.
[0037] Example 4: In an indoor decoration construction site where image sensors are deployed and the ambient light level in the construction area is between 200 lux and 500 lux, the system acquires 100 consecutive video frames of the construction area to complete the benchmark calibration procedure. The spatiotemporal feature processing unit extracts the peak response intensity from the key point heatmap of each frame. Using the formula Calculate the average response intensity benchmark under this physical scenario, where, As the average response intensity benchmark, For the first The peak response intensity of the frame is used by the spatiotemporal feature processing unit to extract the thermal noise amplitude of the image sensor at a preset gain and use it as a scaling factor. This scaling factor is then compared with the average response intensity benchmark. The multiplication generates a preset visibility threshold, which serves as the physical criterion for determining the visibility confidence of key target points.
[0038] When the impact drill operates at a frequency between 20Hz and 50Hz at the construction site, the reference correction module performs a 512-point discrete Fourier transform on the visible anchor point coordinate sequence to extract the energy distribution characteristics of the displacement signal. The system identifies frequency components with energy peaks higher than the limb's spontaneous swing frequency band and determines the frequency starting point of the high-frequency interference zone. The spatiotemporal feature processing unit is based on the relational formula A preset cutoff frequency is determined and set to 4.5Hz. The reference correction module then adjusts the frequency based on this preset cutoff frequency. Configure a Butterworth low-pass filter to perform frequency component separation on the initial coordinate sequence within the historical sampling window, and filter out high-frequency coordinate jump vectors that exceed the preset cutoff frequency. The generated low-frequency reference feature coordinates eliminate coordinate drift caused by the reciprocating motion of the tool and establish a stable spherical center reference point for the spatial constraint module.
[0039] Example 5: In a basement decoration site with insufficient lighting and dust scattering interference during construction, the initial calibration procedure before system deployment establishes the criteria for recognizing stability. The spatiotemporal feature processing unit selects a 7×7 neighborhood pixel set centered on the peak of the thermal image response as a local region, and calculates the arithmetic mean of the response values of each pixel within the local region. Obtain the ambient background noise level and apply the formula. Generate visibility confidence in the current physical scene, where, For visibility confidence, The peak intensity of the thermal response. To determine the average intensity in a local area, the system continuously collects 300 frames of heat map sequences in a static environment without personnel working, and calculates the mean distribution of visibility confidence. 0.6 times the mean distribution is set as the preset visibility threshold, so that the transient pseudo-response component caused by dust particles can be excluded in the subsequent operation phase.
[0040] When construction workers perform plastering work involving significant limb movements and some joints are obscured by building materials piled on site, the reference correction module uses a second-order Butterworth low-pass filter to decouple the visible anchor point coordinate sequence in the frequency domain. To address the time delay deviation caused by fluctuations in the image sensor's acquisition frequency, the system calculates the real-time sampling frequency based on the timestamp interval of consecutive frames and synchronously corrects the filter's transfer function coefficients. The rigid topological radius extracted by the spatial constraint module... Calibration is performed using the median Euclidean distance between target keypoints and visible anchor points within the 10 frames of valid data prior to occlusion. When the coordinate restoration module calculates that the magnitude of the predicted coordinate offset vector in space exceeds the rigid topology radius... At that time, the system projects the coordinate points onto a sphere centered on the low-frequency reference feature coordinates with a radius of [missing information]. On a rigid sphere, the generated spatial feature tensor maintains the conservation of kinematic energy in the time dimension.
[0041] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A monitoring system for interior decoration construction activities, characterized in that, include: The video stream acquisition unit is used to acquire continuous video frames of the construction area; The spatiotemporal feature processing unit, connected to the video stream acquisition unit, is used to extract the coordinate sequence of key points of the human skeleton in continuous video frames and calculate the visibility confidence of the target key points. The spatiotemporal feature processing unit includes: a baseline correction module, used to acquire the coordinate sequence of visible anchor points that have a topological connection with the target key point when the visibility confidence of the target key point is lower than a preset visibility threshold, and to separate the low-frequency displacement components of the visible anchor point coordinate sequence within the historical sampling window, and to generate low-frequency baseline feature coordinates by removing high-frequency noise components with frequencies exceeding a preset cutoff frequency; a spatial constraint module, used to extract the median Euclidean distance between the target key point and the visible anchor points in the historical valid frames to establish the rigid topological radius; and a coordinate repair module, used to calculate the predicted coordinates of the target key point based on kinematic laws, and to calculate the Euclidean distance between the predicted coordinates and the low-frequency baseline feature coordinates; when the Euclidean distance is greater than the rigid topological radius, the vector along the direction vector from the low-frequency baseline feature coordinates to the predicted coordinates is truncated with the rigid topological radius as the scalar length to generate a corrected spatial feature tensor. The action recognition module, connected to the spatiotemporal feature processing unit, is used to calculate the similarity between the spatial feature tensor and the preset template to identify violations.
2. The monitoring system for interior decoration construction activities according to claim 1, characterized in that, The spatiotemporal feature processing unit is used to obtain the centroid spatial acceleration of all visible key points of the human skeleton and adjust the frame length of the historical sampling window according to the magnitude of the centroid spatial acceleration. When the magnitude of the centroid spatial acceleration increases, the spatiotemporal feature processing unit shortens the frame length to reduce the filtering delay. When the magnitude of the centroid spatial acceleration decreases, the spatiotemporal feature processing unit increases the frame length to filter out vibration noise caused by construction tools.
3. The monitoring system for interior decoration construction activities according to claim 1, characterized in that, The spatiotemporal feature processing unit includes a spatiotemporal graph convolutional network, which is used to construct a human topology map using the coordinate sequence of key points of the human skeleton, and to complete the feature flow of occluded key points in situ by using low-frequency reference feature coordinates, so as to eliminate the action semantic discontinuity caused by scaffolding occlusion.
4. A monitoring system for interior decoration construction activities according to claim 1, characterized in that, It also includes an early warning push unit, which is connected to the action recognition module. When a violation is detected, it pushes an early warning signal indicating the construction safety risk to the on-site management terminal. The early warning signal includes the location of the key violation point, the type of violation action, and the current construction work station number.
5. A monitoring system for interior decoration construction activities according to claim 1, characterized in that, The benchmark correction module includes a bypass control branch, which is used to shorten the number of sampling frames in the historical sampling window or stop separating low-frequency displacement components when the displacement intensity of key points of the human skeleton exceeds a preset safety threshold, in order to capture dangerous behaviors such as falls or transient evasion.
6. A monitoring system for interior decoration construction activities according to claim 1, characterized in that, The action recognition module is used to calculate the cosine similarity between the spatial feature tensor and each standard action feature tensor in the preset template, and outputs the preset violation category judgment result when the cosine similarity is greater than the matching threshold.
7. A monitoring system for interior decoration construction activities according to claim 1, characterized in that, The spatiotemporal feature processing unit is used to calculate the coordinates of the geometric center point of the bounding box of the construction tool, and configure the adjacency matrix weights of the local graph convolution operation based on the Euclidean distance between the geometric center point coordinates and the key point at the end of the human skeleton, so as to enhance the feature recognition of the interaction area between the construction tool and the human body.
8. A monitoring system for interior decoration construction activities according to claim 1, characterized in that, The spatiotemporal feature processing unit is used to transform the visible anchor point coordinate sequence to the frequency domain using the Fourier transform operator, and to perform low-pass filtering on the frequency domain signal according to the preset cutoff frequency. The inverse Fourier transform operator is used to restore the filtered signal to the time domain signal in order to extract the low-frequency reference feature coordinates.
9. A monitoring system for interior decoration construction activities according to claim 1, characterized in that, The video stream acquisition unit acquires real-time video data at frame rates of 25Hz to 60Hz from image acquisition devices deployed at the construction site, and uses tracking algorithms to locate the coordinates of key points of the human skeleton.