Method for detecting safety belt hanging state of electric power high-altitude operation personnel based on spatio-temporal context
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HEBEI POWER CONSTR SUPERVISION CO LTD
- Filing Date
- 2025-12-16
- Publication Date
- 2026-05-26
Smart Images
Figure CN122090107A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision, target detection, and power safety production monitoring technology, and in particular to a method and system for detecting the safety belt attachment status of power high-altitude workers based on spatiotemporal context. Background Technology
[0002] Erecting and erecting power transmission towers and routine maintenance are high-risk operations, and workers must strictly adhere to regulations regarding the use of safety harnesses when working at heights. With the development of power Internet of Things (IoT) technology, automated monitoring using climbing robots, drones, or intelligent surveillance cameras has become a trend. However, existing visual inspection solutions suffer from the following main problems in practical engineering applications: Limitations of static detection: Existing technologies mostly use single-stage networks such as YOLO to directly detect the seatbelt wearing status. However, relying solely on a single frame image, when the hook is suspended in front of the tower (not hooked), but visually overlaps with the tower material behind it on the two-dimensional image plane, the algorithm is prone to misjudging it as being hooked. This inability to distinguish between visual overlap and physical anchoring is the main reason for missed detections.
[0003] Interference from dynamic backgrounds: In climbing robot or drone operation scenarios, the camera moves with the equipment, causing both the background (tower) and foreground (hook) in the image to shift. Traditional background subtraction methods or static target detection methods become completely ineffective, making it difficult to determine whether the hook is stationary relative to the tower (true installation) or swaying with the human body (false installation).
[0004] False alarms during operation: It is reasonable for the hook to be in an unattached state when the worker is searching for the attachment point or performing the hooking / unattaching operation. Existing algorithms often lack an understanding of the operation process, and will alarm as soon as they detect that the hook is not attached, resulting in frequent false alarms on site and affecting work efficiency.
[0005] Therefore, a detection method that combines time-dimensional trajectory analysis with spatial-dimensional relative motion logic is needed to accurately determine the seat belt attachment status under moving shooting conditions. Summary of the Invention
[0006] To overcome the shortcomings of the prior art, the present invention provides a method for detecting the safety belt attachment status of power high-altitude workers based on spatiotemporal context, which has the advantages of low computational load and strong resistance to dynamic interference.
[0007] To achieve the above objectives, the present invention adopts the following technical solution, including: A method for detecting the attachment status of safety belts for high-altitude power workers based on spatiotemporal context includes the following steps: S1: Acquire real-time video stream of the high-altitude power operation scene. The video stream is acquired by a moving acquisition device. S2: Build and deploy a detection network to detect each frame of the video stream and output the bounding box coordinates and confidence scores of three types of targets: hook, hanging point, and operator's hand. S3: A multi-target tracking algorithm is used to assign identities and associate trajectories with the detected hooks and attachment points, generating a spatiotemporal trajectory sequence of each target in consecutive frames; S4: Perform preliminary screening in geometric space, calculate the intersection-union ratio of the hook bounding box and the hanging point bounding box, and screen out candidate hanging pairs whose intersection-union ratio meets the spatial overlap threshold. S5: Perform relative motion vector analysis. For candidate anchor pairs, calculate the relative velocity vector of the hook relative to the anchor point based on the spatiotemporal trajectory sequence. When the magnitude of the relative velocity vector is continuously less than the static threshold within the preset time window, the candidate anchor pairs are determined to be in a physical anchoring state. S6: Based on the hand target detection results, the final seat belt attachment state is output using a finite state machine. The seat belt attachment state includes a locked state, a free state, and an operational state.
[0008] Preferably, in step S5, the relative motion vector analysis is specifically as follows: S51: Closest to hooks and hanging points The center point coordinates of the frame are smoothed using a moving average filter. S52, calculate the absolute velocity vector of the hook in the current frame t. And the absolute velocity vector of the hook point (or background feature) in the current frame t. : ; ; Where t represents the current frame; Indicates the set time interval; This represents the coordinates of the center point of the hook after smoothing. This represents the coordinates of the center point of the hanging point after smoothing. S53: Calculate the relative velocity vector of the hook with respect to the attachment point. ; S54: Define the anchorage stability index :
[0009] in: The relative velocity modulus; This is the sensitivity coefficient; This is an indicator function, defined only if the Intersection over Union (IOU) ratio of the hook bounding box and the hook point bounding box is greater than a threshold. Effective immediately, that is ,otherwise ; Set the sliding window duration to N frames, and calculate the average stability index within the window. : ; like If so, the hook is determined to be in a physically anchored state; where To set a stability threshold; like If so, it is determined that the current hook is in a relatively sliding, loose / free state.
[0010] Preferably, in step S52, when the detection network in step S2 does not detect a clear hanging point category, the background velocity vector is calculated using the local optical flow method: The background area is generated by expanding outwards by a preset ratio from the center of the hook bounding box; Corner points are extracted within the background area, and the average displacement vector of the corner points is calculated using an optical flow algorithm, which is then used as the... The alternative value.
[0011] Preferably, in step S51, an exponentially weighted moving average is used for filtering and smoothing: ; in, As a smoothing factor, The coordinates of the center point of the smoothed hook or hanging point bounding box The coordinates of the center of the hook or attachment point bounding box are given, and t represents the current frame.
[0012] Preferably, in step S6, the state transition logic of the finite state machine is as follows: Operation mode: When the overlap ratio (intersection over union ratio) between the bounding box of the operator and the bounding box of the hook is detected to be greater than the operation threshold, the system enters the operation mode and suppresses the alarm signal. Locked state: When in the non-operational state and the physical anchoring state determination condition described in step S5 is met, the system enters the locked state. Free state: When in a non-operational state and the physical anchoring status determination condition described in step S5 is not met for more than a preset confirmation time, the system enters the free state and triggers an alarm.
[0013] Preferably, in step S2, the construction of the detection network is specifically as follows: (1) An improved YOLO architecture is adopted as the base network; (2) Introduce a coordinate attention module after the SPPF module of the backbone network. The specific operation is as follows: First, for the input feature map, coordinate attention is encoded using pooling kernels along the horizontal and vertical directions, respectively, to obtain the output along the horizontal direction and the output along the vertical direction; The horizontal and vertical outputs are then concatenated and processed by a convolutional transformation function to generate an intermediate feature map. Finally, attention weights in the height and width directions are obtained through segmentation and Sigmoid activation, and the original feature map is weighted accordingly. (3) CIoU Loss is used as the bounding box regression loss. The specific calculations are as follows: ; in, Center point of the prediction box Center point of the real frame The square of the Euclidean distance; The diagonal distance of the smallest closure region that can simultaneously contain both the predicted bounding box and the ground truth bounding box; These are weight parameters; Used to measure the consistency of aspect ratio.
[0014] Preferably, in step S3, the ByteTrack algorithm is used to assign the identity of the hook and the target at the hook point and associate the trajectory.
[0015] This invention also provides a spatiotemporal context-based safety belt attachment status detection system for power high-altitude workers, applicable to the aforementioned spatiotemporal context-based safety belt attachment status detection method for power high-altitude workers. The system includes: Data acquisition module: includes a high-definition camera mounted on a climbing robot or drone to acquire real-time video streams containing power workers at height and the environment; Edge computing module: An embedded processor deployed with a detection network and logic operations; Results output module: used to visualize the detection results and trigger audible and visual alarms or remote push notifications when a violation is detected.
[0016] The present invention also provides an electronic device, which includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the aforementioned method for detecting the attachment status of safety belts for high-altitude power workers based on spatiotemporal context.
[0017] The present invention also provides a computer program product, which includes a computer program / instruction that, when executed by a processor, implements the aforementioned method for detecting the attachment status of safety belts for high-altitude power workers based on spatiotemporal context.
[0018] The advantages of this invention are: (1) Unlike traditional static image analysis, this invention completely eliminates the background displacement interference caused by the movement / vibration of the climbing robot or drone itself through relative velocity vector analysis, and can accurately identify the real and fake hanging devices, with strong anti-dynamic interference capability.
[0019] (2) This invention is only applicable when A device is only considered physically mounted when its IoU consistently approaches zero (relatively static) over a period of time and meets a threshold. If the hook is too large, it means that although it overlaps with the tower in the picture, it is actually moving with the human body or in a sliding state, and is judged as a false hook.
[0020] (3) In certain extreme perspectives (such as close-up shots), the detection network may fail to identify the complete hanging point category. In this case, it automatically downgrades to optical flow-assisted mode, using the local optical flow method to calculate the background velocity vector as... The alternative value. This mechanism ensures that even when a clear anchor cannot be detected, the environmental texture can still be used to determine whether the hook is suspended.
[0021] (4) This invention abandons the semantic segmentation or large model scheme with high computing power requirements, and only relies on lightweight target detection and algebraic operation. The detection speed is fast, the algorithm is lightweight and efficient, and it is suitable for deployment in the low-power embedded computing unit of climbing robot.
[0022] (5) This invention introduces a hand detection and state machine mechanism, which conforms to the actual operating habits of workers, effectively avoids invalid alarms during operation, has rigorous logic, low false alarm rate, and improves the practicality of the system. The detection network of this invention adopts an improved YOLO series (such as YOLOv8-Nano) as the base model to meet the real-time requirements of embedded edge devices, and is a lightweight end-to-end detection network.
[0023] (6) The detection network of the present invention introduces a coordinate attention module in the backbone network to enhance the network’s sensitivity to the slender structure and orientation information of the hook, suppress the interference of complex tower material background, and enable the network to accurately locate the hook head and hook body position. In the background of cluttered iron tower texture, the recall rate of the hook target is increased by about 12%.
[0024] (7) The detection network of the present invention uses CIoU Loss to optimize the localization loss, which effectively solves the problem of gradient vanishing when the predicted box and the real box do not overlap, so that the model can still converge when the hook is tilted at a large angle, and improves the regression accuracy of small target bounding boxes.
[0025] (8) This invention has the advantages of small computational load and strong anti-dynamic interference capability. It is suitable for deployment on edge devices with limited computing power. Through lightweight detection and relative motion vector analysis, it can realize high-precision state perception on edge devices with low computing power. Attached Figure Description
[0026] Figure 1 This is a flowchart of the method for detecting the safety belt attachment status of high-altitude power workers based on spatiotemporal context, according to the present invention.
[0027] Figure 2 This is a schematic diagram of "real installation - physical anchoring" and "fake installation - visual overlap but suspended".
[0028] Figure 3 This is an architecture diagram of the power high-altitude operation personnel safety belt attachment status detection system based on spatiotemporal context according to the present invention. Detailed Implementation
[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0030] Example 1 Depend on Figure 1 As shown, the method for detecting the safety belt attachment status of power workers at heights based on spatiotemporal context includes the following steps: S1: Acquire real-time video stream of the high-altitude power operation scene. The video stream is acquired by a moving acquisition device.
[0031] In this embodiment, construction images under different lighting conditions (sunny, cloudy, backlight) and different tower types were collected to construct a dataset. Hooks, anchor points (such as angle iron and bolts), and hands in the images were classified and labeled, with multi-scale data augmentation performed specifically for small hooks. The data acquisition equipment consisted of a climbing robot or a drone.
[0032] S2: Build and deploy a lightweight end-to-end detection network to detect each frame of the video stream and output the bounding box coordinates and confidence scores of three types of targets: hook, anchor, and hand.
[0033] Considering the limited computing power of embedded computing units (such as NVIDIA Jetson Nano or Rockchip RK3588) carried by climbing robots or drones, this embodiment uses an improved YOLOv8-Nano architecture as the base network and makes the following improvements: (1) Improvement of the network backbone (Coordinate Attention embedding) Traditional convolutional neural networks often focus only on local information and ignore positional information when extracting features. Since safety hooks are usually long and thin strips and are easily confused in complex backgrounds (such as interlaced tower materials), this embodiment embeds a coordinate attention (CA) mechanism after the SPPF (Spatial Pyramid Pooling - Fast) module of the backbone network.
[0034] The specific calculation process is as follows: First, regarding the input feature map (size is) The CA module uses dimensions along the horizontal (X-axis) and vertical (Y-axis) directions respectively. and Encoding is performed using pooling kernels; Output along the X-axis : ; Output along the Y-axis : ; The above operation aggregates features from two spatial directions to generate a pair of direction-aware feature maps.
[0035] Then the two are concatenated (Concat), after... Convolution Transform Function Generate intermediate feature maps : ;in, It is a non-linear activation function.
[0036] Finally, attention weights in the height and width directions are obtained through segmentation and Sigmoid activation. and And weight the original feature map: .
[0037] This improvement enabled the network to accurately locate the hook head and body of the hook, increasing the recall rate of the hook target by approximately 12% in a cluttered background of Eiffel Tower textures.
[0038] (2) Optimization of the loss function (CIoU Loss) To address the issue of hooked targets being extremely small objects with drastically varying aspect ratios (due to different shooting angles), this embodiment abandons the traditional IoU Loss and adopts CIoU Loss (Complete IoU Loss) as the bounding box regression loss: ; in: The squared Euclidean distance between the center point of the predicted bounding box and the center point of the true bounding box; The diagonal distance of the smallest closure region that can simultaneously contain both the predicted bounding box and the ground truth bounding box; These are weight parameters; Used to measure the consistency of aspect ratio: ; CIoU Loss effectively solves the gradient vanishing problem when the predicted bounding box and the ground truth bounding box do not overlap, enabling the model to converge even when the hook is tilted at a large angle.
[0039] S3: A multi-target tracking algorithm is used to assign identities (IDs) and trajectories to the detected hooks and attachment points, generating a spatiotemporal trajectory sequence of each target in consecutive frames.
[0040] S4: Perform preliminary screening in geometric space, calculate the intersection-over-union ratio (IoU) of the hook bounding box and the hanging point bounding box, and screen out candidate hanging pairs whose IoU meets the spatial overlap threshold; if the IoU does not meet the spatial overlap threshold, it means that the hook and the hanging point are not connected.
[0041] Because the camera is in motion and the hook may be briefly obstructed by the operator's arm, a single detection cannot maintain ID consistency. This embodiment uses the ByteTrack algorithm to assign the identity (ID) of the hook and the target at the attachment point and associate it with the trajectory.
[0042] The ByteTrack algorithm is shown below: (1) Definition of state vector For each detected hook or attachment point, a Kalman filter state vector is established. : ;in The coordinates of the bounding box center are For area, Aspect ratio; These represent the corresponding rates of change.
[0043] (2) Cascaded matching strategy First-level matching (high-scoring boxes): Confidence level... The high-resolution detection bounding boxes are matched with existing tracks using IoU.
[0044] Second-level matching (low-scoring boxes): Matching unmatched trajectories with confidence levels... The low-scoring detection boxes between the two are then matched a second time. This step is crucial because when the hook moves rapidly and produces motion blur, the detection confidence decreases. The ByteTrack algorithm can effectively recover these blurred targets, ensuring that the trajectory does not break.
[0045] S5: Perform relative motion vector analysis. For candidate hook-and-hook pairs, calculate the relative velocity vector of the hook relative to the hook point based on the spatiotemporal trajectory sequence. When the magnitude of the relative velocity vector is continuously less than the static threshold within a preset time window, determine that the candidate hook-and-hook pair is in a physical anchoring state.
[0046] To determine physical anchorage in a dynamic context, this invention designs a relative motion vector analysis algorithm, as detailed below: S51: For the hooked target and the anchor point target in the most recent The center point coordinates of the frame are smoothed using a moving average filter.
[0047] In this embodiment, the specific method for trajectory smoothing (denoising) is as follows: the center coordinates of the detection box (hook or hanging point bounding box) This is usually accompanied by pixel-level jitter, which is smoothed using an exponentially weighted moving average (EWMA): ;in, As a smoothing factor, The coordinates of the smoothed center point are given, and t represents the current frame.
[0048] S52: Calculate the absolute velocity vector of the hook target in the current frame t. And the absolute velocity vector of the anchor (or background feature) in the current frame t. .
[0049] In this embodiment, the central difference method is used to improve robustness: ; ; Where t represents the current frame time; Indicates the set time interval; Furthermore, under certain extreme viewing angles (such as close-up shots), the detection network in step S2 may fail to identify the complete angle steel or hanging point category. In this case, it automatically downgrades to optical flow-assisted mode, using the local optical flow method to calculate the background velocity vector: (1) Feature point generation: Taking the hook prediction box (bounding box) as the center, within a region that extends outward by 2.5 times from the center point of the hook prediction box, the foreground region belonging to the hook and the human body is removed, and only the background region is retained; (2) Corner extraction: K corners (StrongCorners) are extracted from the background area using the Shi-Tomasi algorithm. (3) LK optical flow tracing: The Lucas-Kanade sparse optical flow algorithm is used to calculate The position of each corner point in the next frame; (4) Background velocity estimation: The RANSAC (Random Sample Consensus) algorithm is used to remove outliers, and the average displacement vector of the remaining corner points is calculated to obtain the background velocity. The estimated value, as the stated The alternative value.
[0050] This mechanism ensures that even when a clear anchor cannot be detected, environmental textures can still be used to determine whether a hook is suspended.
[0051] S53: Calculate the relative velocity vector of the hook with respect to the attachment point. ; like Figure 2 As shown, only when A device is only considered physically mounted when its IoU consistently approaches zero (relatively static) over a period of time and meets a threshold. If the hook is too large, it means that although it overlaps with the tower in the picture, it is actually moving with the human body or in a sliding state, and is judged as a false hook.
[0052] S54: Define the Anchoring Stability Index (ASI) : ; in: The relative velocity modulus; This is the sensitivity coefficient (an empirical value of 5.0). This is an indicator function, applied only when the overlap between the hook and the hanging point is greater than a threshold. Effective immediately, that is ,otherwise .
[0053] This invention maintains a length of (like A sliding window (frames) is used to calculate the average stability index within the window. : ; like If the hook is physically anchored, it means that the hook and the attachment point overlap in space and remain relatively stationary. like If the hook is in a relatively sliding, loose / free state, it means that although there may be overlap, there is relative slippage or swaying between the two.
[0054] S6: Based on the hand target detection results, the final seat belt attachment state is output using a finite state machine. The seat belt attachment state includes locked state (safe), free state (violation), and operating state (operating).
[0055] To conform to workers' operating habits and avoid interfering with normal operations, this embodiment designs a finite state machine (FSM) with four states, as shown below: (1) State definition A0: No overlap between hook and hanging point detected.
[0056] A1: The hand is gripping the hook.
[0057] A2: The hand leaves, the hook and the hanging point overlap, and the timing determination of N frames is in progress.
[0058] A3: Detection successful, locked to safe status.
[0059] A4: Judgment failed (timeout or slippage), locked as a violation status, and an alarm is triggered.
[0060] (2) State transition conditions A0 → A1: When the overlap between the operator's target bounding box and the hook's target bounding box, i.e., the Intersection over Union (IoU) (Hand, Hook), is detected to be greater than the operation threshold, i.e., IoU (Hand, Hook) > 0.3, the system becomes silent, enters the operation state, and suppresses alarm signals.
[0061] A1->A2: Detected IoU(Hand,Hook) < 0.1 (hand removed). System starts timer. .
[0062] A2—>A3: Timer The end, and At this point, the system enters a locked state and displays a green safety box.
[0063] A2 → A4: Timer The end, and Or during the timing process The system returns to zero. At this point, the system enters a free state and triggers an alarm, displaying a red alarm box and activating the buzzer.
[0064] A3 → A0: Continuous If the frame detects that IoU(Hook,Anchor) < 0.1, it means that the hook has been removed.
[0065] Example 2 Depend on Figure 3 As shown, the spatiotemporal context-based safety belt attachment status detection system for high-altitude power workers is applicable to the spatiotemporal context-based safety belt attachment status detection method for high-altitude power workers in Embodiment 1 above, including: Data acquisition module: includes a high-definition camera mounted on the climbing robot or drone to capture video streams containing personnel and the environment; Edge computing module: Deployed with an embedded processor capable of executing the detection network and logic operations of Embodiment 1 above; The results output module is used to visualize the detection results and trigger an audible and visual alarm or remote push when the state is determined to be in violation (free state).
[0066] Example 3 An electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method for detecting the safety belt attachment status of high-altitude power workers based on spatiotemporal context as described in Embodiment 1 above.
[0067] The electronic device in this application embodiment may be the mobile device itself, or a standalone device independent of it. The standalone device may communicate with the mobile device to receive the collected input signals from it and send the selected target decision behavior to it.
[0068] An electronic device includes one or more processors and memory. The processor may be a central processing unit (CPU) or other processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device to perform desired functions. The memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor may execute the program instructions to implement the decision-making behavior and decision-making methods of the various embodiments of this application described above, and / or other desired functions.
[0069] Electronic devices may also include input devices and output devices.
[0070] Example 4 In addition to the methods and devices described above, embodiments of this application may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps in the decision-making behavior decision-making method according to various embodiments of this application as described in Embodiment 1 above.
[0071] The computer program product can be written in any combination of one or more programming languages to perform the operations of the embodiments of this application. The programming languages include object-oriented programming languages such as Java and C++, as well as conventional procedural programming languages such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.
[0072] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for detecting the safety belt attachment status of power workers at heights based on spatiotemporal context, characterized in that, Includes the following steps: S1: Acquire real-time video stream of the high-altitude power operation scene. The video stream is acquired by a moving acquisition device. S2: Build and deploy a detection network to detect each frame of the video stream and output the bounding box coordinates and confidence scores of three types of targets: hook, hanging point, and operator's hand. S3: A multi-target tracking algorithm is used to assign identities and associate trajectories with the detected hooks and attachment points, generating a spatiotemporal trajectory sequence of each target in consecutive frames; S4: Perform preliminary screening in geometric space, calculate the intersection-union ratio of the hook bounding box and the hanging point bounding box, and screen out candidate hanging pairs whose intersection-union ratio meets the spatial overlap threshold. S5: Perform relative motion vector analysis. For candidate anchor pairs, calculate the relative velocity vector of the hook relative to the anchor point based on the spatiotemporal trajectory sequence. When the magnitude of the relative velocity vector is continuously less than the static threshold within the preset time window, the candidate anchor pairs are determined to be in a physical anchoring state. S6: Based on the hand target detection results, the final seat belt attachment state is output using a finite state machine. The seat belt attachment state includes a locked state, a free state, and an operational state.
2. The method for detecting the safety belt attachment status of high-altitude power workers based on spatiotemporal context according to claim 1, characterized in that, In step S5, the relative motion vector analysis is as follows: S51: Closest to hooks and hanging points The center point coordinates of the frame are smoothed using a moving average filter. S52, calculate the absolute velocity vector of the hook in the current frame t. And the absolute velocity vector of the hook point (or background feature) in the current frame t. : Where t represents the current frame; Indicates the set time interval; This represents the coordinates of the center point of the hook after smoothing. This represents the coordinates of the center point of the hanging point after smoothing. S53: Calculate the relative velocity vector of the hook with respect to the attachment point. ; S54: Define the anchorage stability index : in: The relative velocity modulus; This is the sensitivity coefficient; This is an indicator function, defined only if the Intersection over Union (IOU) ratio of the hook bounding box and the hook point bounding box is greater than a threshold. Effective immediately, that is ,otherwise ; Set the sliding window duration to N frames, and calculate the average stability index within the window. : like If so, the hook is determined to be in a physically anchored state; where To set a stability threshold; like If so, it is determined that the current hook is in a relatively sliding, loose / free state.
3. The method for detecting the safety belt attachment status of power workers at heights based on spatiotemporal context according to claim 2, characterized in that, In step S52, if the detection network in step S2 does not detect a clear hanging point category, the background velocity vector is calculated using the local optical flow method: The background area is generated by expanding outwards by a preset ratio from the center of the hook bounding box; Corner points are extracted within the background area, and the average displacement vector of the corner points is calculated using an optical flow algorithm, which is then used as the... The alternative value.
4. The method for detecting the safety belt attachment status of high-altitude power workers based on spatiotemporal context according to claim 2, characterized in that, In step S51, an exponentially weighted moving average is used for filtering and smoothing: ;in, As a smoothing factor, The coordinates of the center point of the smoothed hook or hanging point bounding box The coordinates of the center of the hook or attachment point bounding box are given, and t represents the current frame.
5. The method for detecting the safety belt attachment status of power workers at heights based on spatiotemporal context according to claim 1, characterized in that, In step S6, the state transition logic of the finite state machine is as follows: Operation mode: When the overlap ratio (intersection over union ratio) between the bounding box of the operator and the bounding box of the hook is detected to be greater than the operation threshold, the system enters the operation mode and suppresses the alarm signal. Locked state: When in the non-operational state and the physical anchoring state determination condition described in step S5 is met, the system enters the locked state. Free state: When in a non-operational state and the physical anchoring status determination condition described in step S5 is not met for more than a preset confirmation time, the system enters the free state and triggers an alarm.
6. The method for detecting the safety belt attachment status of power workers at heights based on spatiotemporal context according to claim 1, characterized in that, In step S2, the construction of the detection network is specifically as follows: (1) An improved YOLO architecture is adopted as the base network; (2) Introduce a coordinate attention module after the SPPF module of the backbone network. The specific operation is as follows: First, for the input feature map, coordinate attention is encoded using pooling kernels along the horizontal and vertical directions, respectively, to obtain the output along the horizontal direction and the output along the vertical direction; The horizontal and vertical outputs are then concatenated and processed by a convolutional transformation function to generate an intermediate feature map. Finally, attention weights in the height and width directions are obtained through segmentation and Sigmoid activation, and the original feature map is weighted accordingly. (3) CIoU Loss is used as the bounding box regression loss. The specific calculations are as follows: in, Center point of the prediction box Center point of the real frame The square of the Euclidean distance; The diagonal distance of the smallest closure region that can simultaneously contain both the predicted bounding box and the ground truth bounding box; These are weight parameters; Used to measure the consistency of aspect ratio.
7. The method for detecting the safety belt attachment status of power workers at heights based on spatiotemporal context according to claim 1, characterized in that, In step S3, the ByteTrack algorithm is used to assign the identities and associate the trajectories of the hook and the target.
8. A power high-altitude work personnel safety belt attachment status detection system based on spatiotemporal context, characterized in that, The system, applicable to the spatiotemporal context-based safety belt attachment status detection method for high-altitude power workers as described in any of claims 1-7, comprises: Data acquisition module: includes a high-definition camera mounted on a climbing robot or drone to acquire real-time video streams containing power workers at height and the environment; Edge computing module: An embedded processor deployed with a detection network and logic operations; Results output module: used to visualize the detection results and trigger audible and visual alarms or remote push notifications when a violation is detected.
9. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method for detecting the attachment status of safety belts for high-altitude power workers based on spatiotemporal context as described in any one of claims 1-7.
10. A computer program product, characterized in that, It includes a computer program / instruction that, when executed by a processor, implements the method for detecting the attachment status of safety belts for high-altitude power workers based on spatiotemporal context as described in any one of claims 1-7.