A standard penetration test measurement method based on machine vision

Abstract

The application provides a kind of standard penetration test measurement method based on machine vision.The measurement method is to install a height reference rod vertically beside the drill rod of standard penetration test, and to set up a high-definition monocular camera at a distance of 2.5-3.5 meters from the test device to record the whole process video of the test;target detection is carried out frame by frame for all videos in the test process to form the original height sequence of the hammer body at different times in the standard penetration test process, and Gaussian smoothing processing is carried out on it;the sliding window difference method is used to calculate the motion trend to judge the state of the hammer body;the starting position coordinates and the ending position coordinates of each hammering are obtained from the original height sequence of the hammer body, the difference between the two is calculated as the pixel displacement corresponding to the current hammering, and the actual penetration depth of each hammering is converted.The application realizes the motion state recognition and the number of blows of standard penetration hammer through machine vision technology, significantly improves the operation efficiency, and can obtain reliable scale factor and strong anti-interference ability.

Description

Technical Field

[0001] This invention relates to the field of civil engineering surveying technology, specifically a standard penetration test measurement method based on machine vision. This method uses machine vision technology to automatically record the number of hammer blows and the penetration depth in a standard penetration test. Background Technology

[0002] Currently, the measurement scheme used in the actual operation of the Standard Penetration Test (SPT) is as follows: First, the staff will carry out preliminary preparation work, namely, a 15 cm pre-driving operation. This step is to ensure the smooth conduct of the subsequent formal test. After the pre-driving work is completed, the preparation stage before the formal test begins. At this time, the staff will use a steel tape measure and, with the aid of tape or chalk, mark specific positions of 10 cm, 20 cm, and 30 cm on the drill rod. The purpose of these markings is to record the number of hammer blows within each 10 cm interval. However, the accuracy of the recording method in the existing measurement method largely depends on the manual measurement and recording process. Due to the manual operation, there is inevitably a certain degree of subjectivity and error. In actual operation, some workers may choose to estimate the number of hammer blows by visual inspection or based on personal experience for convenience or other factors, rather than strictly following the markings for precise measurement. This practice obviously lacks an effective supervision mechanism and is very likely to result in inaccurate test data, or even significant deviations.

[0003] However, both domestic and international standards, such as my country's current "Code for Geotechnical Investigation" GB50021, clearly stipulate that the specific data of penetration depth must be accurately recorded during standard penetration tests. Given the numerous shortcomings and potential risks of the aforementioned manual measurement methods, it is essential to introduce advanced recording and measurement methods based on machine vision technology to ensure the accuracy and reliability of test data. Machine vision technology enables automated and precise recording and measurement of the standard penetration test process, thereby effectively improving the scientific validity and credibility of the test data. Summary of the Invention

[0004] To address the problems existing in the background technology, this invention proposes a standard penetration test measurement method based on machine vision. This method introduces a height reference object with known physical dimensions at the test site, and combines visual tracking of the hammer's movement trajectory with intelligent state recognition to accurately convert the pixel displacement observed in the image into the actual physical distance, thereby achieving fully automatic and high-precision recording of the number of hammer blows and the penetration depth.

[0005] To achieve the above technical objectives, the present invention provides a standard penetration test measurement method based on machine vision, characterized in that the steps of the measurement method are as follows:

[0006] S1. Install a height reference rod vertically next to the drill pipe of the standard penetration test. The height reference rod and the standard penetration hammer are on the same plane. Set up a high-resolution monocular camera 2.5 to 3.5 meters away from the test device to ensure that the hammer and the reference rod are in the center of the frame and unobstructed throughout the entire movement. Record the entire process of the standard penetration test with the camera.

[0007] S2. Perform target detection frame-by-frame on all videos of the standard penetration test process to obtain the vertical pixel coordinates of the bottom center point of the SPT hammer bounding box at each moment during the standard penetration test. This constitutes a sequence of the original height of the hammer at different times during the standard penetration test. ;

[0008] S3. The original height sequence of the hammer in step S2. Each value in Gaussian smoothing is performed to eliminate high-frequency noise, resulting in a smoothed hammer height sequence. And the vertical pixel coordinates of the bottom center point of each standard penetration test hammer bounding box. The processing procedure is as follows:

[0009] in, The vertical pixel value of the bottom center point of each SPT hammer bounding box after Gaussian smoothing; yes The original height of the hammer at time t; 'i' is a summation index representing an offset of the point coordinates around time t, with a value ranging from -N to N. The data from the preceding and following N points are used for weighted summation; Standard deviation Gaussian kernel; The kernel half-width is set to 30.

[0010] S4. The motion trend is calculated using the sliding window difference method; the time window length is defined as a value 'a', which is between one-quarter and one-half of the shooting frame rate, and the smoothed hammer height difference between the current frame and the frame a before is calculated. :

[0011] in, It is the height difference between the smoothed hammer height in frame t and the smoothed hammer height in frame a. It is the hammer height value in the t-th frame of the smooth sequence; In smooth sequences Frame hammer height value;

[0012] according to The sign and amplitude of the hammer are combined with the current state of the hammer to dynamically determine the hammer's motion state, which includes falling, stationary, and rising.

[0013] S5. When the state changes from falling to still and the minimum stillness duration is met, a valid hammer strike is determined to have occurred, and the current frame index is recorded as the end time of this hammer strike. The time when the last hammer blow ended is recorded as the start time of the current hammer blow. ;

[0014] S6. Using the original hammer height sequence from step S1 The vertical pixel coordinates of the bottom center point of the standard penetration test hammer bounding box at the end of the last hammer blow are obtained as the coordinates of the starting position of the current hammer blow. And obtain the end time of this hammer strike. The vertical pixel coordinates of the bottom center point of the corresponding standard penetration test hammer bounding box are used as the coordinates of the end position of this hammer strike. The difference between the two values ​​is calculated as the pixel displacement corresponding to this hammer strike:

[0015]

[0016] At the same time, the end time of this hammering will be... Update to the start time of the next hammer strike. This serves as the starting point for calculating the displacement of the next hammer blow;

[0017] S7. Scale calibration and depth conversion; uniformly select at least 10 frames from the video that contain complete reference rods, and calculate the total pixel height of the reference rods in each frame. Frames that deviate from the mean by more than two standard deviations are removed, and the average value of the remaining frames is taken. Assuming the final pixel height; let the total physical height of the reference rod be... If the unit is cm, then the pixel-physical conversion factor is:

[0018] The unit is cm / pixel;

[0019] S8. Calculate the actual penetration depth for each hammer blow. for: ;

[0020] System cumulative When the cumulative depth reaches 10 cm, 20 cm, and 30 cm respectively, the corresponding number of hammer blows is recorded to generate standard penetration test result data.

[0021] A further technical solution of the present invention: In step S2, a deep learning target detection model is used to simultaneously identify the standard penetration test hammer body and the dedicated height reference rod, obtain the lower left corner coordinates of the standard penetration test hammer bounding box, extract its vertical Y-axis position from it, and form a hammer height sequence that changes over time; the target detection model is trained using the YOLO model or the DETR model as the framework model.

[0022] The preferred technical solution of this invention is as follows: In step S1, the camera records the entire experimental process video at a frame rate of 30 frames per second; in step S4, the time window length is defined as 15 frames, and the height difference between the current frame and 15 frames ago is calculated. :

[0023] .

[0024] The preferred technical solution of the present invention is as follows: the height reference rod in step S1 is composed of two equal-length red and green rods, each rod being 60 cm high, for a total height of 120 cm.

[0025] The preferred technical solution of this invention: The hammer state judgment rule in step S4 is as follows:

[0026] Set the threshold based on the height reference pole pixel height obtained in step S7. and ;

[0027] When in a "stationary" state:

[0028] like The hammer body is determined to begin rising.

[0029] like and The hammer is determined to have started falling.

[0030] When in an "ascending" state:

[0031] like This indicates that the upward trend has weakened or stopped, and the market has returned to a static state.

[0032] If both conditions are met and If so, it is determined that the hammer body has changed from rising to falling.

[0033] When in a "falling" state:

[0034] like If the descent is considered to be nearing a stop, a timer is started; if this condition is met continuously for at least If a frame is reached, it is determined that the hammer has stabilized and stopped, and a valid hammer strike record is triggered.

[0035] like This indicates that the hammer bounces back up immediately after falling, and the state switches to rising.

[0036] The preferred technical solution of the present invention is that the minimum static duration in step S5 can determine the time interval between the hammer falling and coming to rest, and is less than the time interval between the worker's operation of pulling it up again, and is set to 0.1s to 0.5s.

[0037] The preferred technical solution of the present invention is as follows: In step S7, the existing HSV color segmentation and extraction algorithm is used to calculate the total pixel height of the reference rod in each frame.

[0038] The preferred technical solution of this invention is as follows: The target detection model adopts a target detection model based on the YOLOv8-small architecture, and its training process is as follows: First, image and video data containing the standard penetration test hammer, height reference rod and their working environment are collected at multiple standard penetration test sites, covering different lighting conditions, background complexity and hammer movement state; and the two types of targets in the images are manually labeled with annotation tools, namely "standard penetration hammer" and "height reference rod", forming an annotated training dataset;

[0039] Subsequently, a transfer learning strategy was adopted, using the YOLOv8-small model pre-trained on a general image dataset as the initial weights, to fine-tune the training on the aforementioned special dataset; and data augmentation methods were introduced during the training process to improve the model's adaptability to actual working conditions; the data augmentation methods included image scaling, color perturbation, and geometric transformation.

[0040] The model training is based on the detection accuracy on the validation set as the convergence criterion. Training is terminated when the performance tends to stabilize, thus obtaining a deep learning object detection model.

[0041] The present invention has the following significant advantages and beneficial effects:

[0042] (1) This invention uses machine vision technology to realize the recognition of the motion state of the standard penetration hammer and the automatic counting of blows, completely eliminating the dependence on manual visual inspection and recording, significantly improving work efficiency, and achieving a high degree of automation;

[0043] (2) This invention supports post-event playback verification by recording video and structured data throughout the process, enhances the transparency and supervision of the test process, meets the requirements for quality control of engineering survey data, and has strong data traceability;

[0044] (3) This invention only requires a common industrial camera and a lightweight edge computing device (such as Jetson Nano or industrial computer) to be quickly deployed on site without structural modification of existing drilling equipment. It has good engineering applicability and promotion value, strong adaptability and convenient deployment.

[0045] (4) This invention uses a dual-path mechanism of “smoothing sequence for state judgment + original sequence for displacement calculation” to suppress noise interference and avoid depth underestimation caused by filtering, thus ensuring the accuracy of physical quantity conversion. It adopts a multi-frame fusion and statistical anomaly removal strategy to significantly improve the robustness of reference rod height measurement. Even if some frames have temporary occlusion, reliable scale coefficients can still be obtained, and the anti-interference ability is strong. Attached Figure Description

[0046] Figure 1 This is an overall flowchart of the present invention;

[0047] Figure 2 This is a schematic diagram of the state transition of the penetrating hammer in this invention;

[0048] Figure 3 This is a layout diagram of the experimental apparatus of the present invention. Detailed Implementation

[0049] The present invention will be further described below with reference to specific embodiments.

[0050] In this embodiment, during the standard penetration test (SPT), a monocular industrial camera continuously captures video of the test area. A deep learning target detection model (such as YOLO or DETR) deployed on field terminal equipment simultaneously identifies two key targets: the SPT hammer itself and a dedicated height reference rod. After specialized training, the model can stably output the coordinates of the lower left corner of the SPT hammer's bounding box, extracting its vertical (Y-axis) position to form a time-varying hammer height sequence.

[0051] In this embodiment, the target detection model adopts a target detection model based on the YOLOv8-small architecture. Its training process is as follows: First, image and video data containing the standard penetration test hammer, height reference pole, and their working environment are collected at multiple standard penetration test sites, covering different lighting conditions, background complexity, and hammer movement states. Two types of targets in the images are manually labeled using annotation tools: "standard penetration hammer" and "height reference pole," forming an annotated training dataset. Then, a transfer learning strategy is adopted, using the YOLOv8-small model pre-trained on a general image dataset as initial weights, to fine-tune the aforementioned specialized dataset. Data augmentation methods are introduced during training to improve the model's adaptability to actual working conditions. These data augmentation methods include image scaling, color perturbation, and geometric transformation. The model training uses the detection accuracy on the validation set as the convergence criterion. Training is terminated when the performance stabilizes, resulting in a deep learning target detection model.

[0052] The embodiment provides a standard penetration test measurement method based on machine vision, the overall process of which is as follows: Figure 1 As shown, the specific steps are as follows: S1. On-site setup and image acquisition, as detailed below. Figure 3 As shown, a height reference rod is vertically installed next to the standard penetration test drill rod, ensuring that the height reference rod is on the same plane as the standard penetration hammer to reduce conversion errors. The rod consists of two equal-length red and green blocks, each 60 cm high, for a total height of 120 cm. A high-resolution monocular camera is set up 2.5 to 3.5 meters away from the test setup, ensuring that the hammer and the reference rod are centered in the frame without obstruction throughout the entire movement. The camera records the entire test process at a frame rate of 30 frames per second.

[0053] S2. Perform target detection frame-by-frame on all videos of the standard penetration test process to obtain the vertical pixel coordinates of the bottom center point of the SPT hammer bounding box at each moment during the standard penetration test. This constitutes a sequence of the original height of the hammer at different times during the standard penetration test. The pixel height here is defined with the top left corner of the video frame as the origin (0,0), the positive X-axis direction to the right, and the positive Y-axis direction downwards. The Y-axis pixel coordinates of the bottom center point of the bounding box are detected. There are multiple hammer blows in the standard penetration test process, and the process of multiple hammer blows needs to be recorded here. In the embodiment, t can be understood as time or frame.

[0054] Assuming the video shooting frame rate is 30 frames per second, taking a 60-second video as an example, then t = 1, 2, ..., 1800;

[0055] y1 = 1 / 30 second (frame 1) Y-axis coordinate of the bottom center of the time-scaled penetrating hammer detection frame

[0056] y2 = 2 / 30 seconds (second frame) Y-axis coordinate of the bottom center of the time-stamped penetrating hammer detection frame ...

[0058] y30 = 1 second (30th frame) Y-axis coordinate of the bottom center of the penetrating hammer detection frame ...

[0060] y1800 = the Y-axis coordinate of the bottom center of the penetrating hammer detection frame at 60 seconds (1800th frame).

[0061] S3. The original height sequence of the hammer in step S2. Each value in Gaussian smoothing is performed to eliminate high-frequency noise, resulting in a smoothed hammer height sequence. And the vertical pixel coordinates of the bottom center point of each standard penetration test hammer bounding box. The processing procedure is as follows:

[0062]

[0063] in, The vertical pixel value of the bottom center point of each SPT hammer bounding box after Gaussian smoothing; yes The original height of the hammer at time t; 'i' is a summation index representing an offset of the point coordinates around time t, with a value ranging from -N to N. The data from the preceding and following N points are used for weighted summation; Standard deviation Gaussian kernel; The kernel half-width is set to 30.

[0064] The calculation here requires the current point. and its before and after points (i.e., from) arrive ) participates in the weighted average, here You can refer to the video frame rate and choose a suitable value. In this example, we set the kernel half-width to 30.

[0065] S4. Calculate motion trend using the sliding window difference method; define the time window length as one-quarter to one-half of the shooting frame rate; given a shooting frame rate of 30 frames / second, define the time window length as 15 frames, leaving an appropriate time window to determine if displacement has occurred; calculate the height difference between the current frame and 15 frames ago. :

[0066] ;

[0067] in, It is the height difference between the smoothed hammer height in frame t and the smoothed hammer height 15 frames ago;

[0068] according to The sign and amplitude of the hammer are used, combined with the current state of the hammer, to dynamically determine the hammer's motion state, which includes falling, stationary, and rising states. The rules for determining the hammer's state are as follows:

[0069] threshold and The setting range can be referenced to the pixel height of the height reference rod obtained from S7; assuming the obtained reference rod pixel height is 600 pixels, a movement of 2cm up or down is considered movement, and this range can be set accordingly. 2cm / 120cm * 600 pixels = 10 pixels The settings are the same;

[0070] When in a "stationary" state:

[0071] like (like (pixels), determining that the hammer has started to rise;

[0072] like and (like (pixels / 15 frames), indicating that the hammer has started to fall;

[0073] When in an "ascending" state:

[0074] like This indicates that the upward trend has weakened or stopped, and the market has returned to a static state.

[0075] If both conditions are met and If so, it is determined that the hammer body has changed from rising to falling.

[0076] When in a "falling" state:

[0077] like If the descent is considered to be nearing a stop, a timer is started; if this condition is met continuously for at least If a frame is reached, it is determined that the hammer has stabilized and stopped, and a valid hammer strike record is triggered.

[0078] like This indicates that the hammer bounces back up immediately after falling, and the state switches to rising.

[0079] It is the hammer height value in the t-th frame of the smooth sequence; In smooth sequences Frame hammer height value.

[0080] S5. When the state changes from falling to still and the minimum stillness duration is met, a valid hammer strike is determined to have occurred, and the current frame index is recorded as the end time of this hammer strike. The time when the last hammer blow ended is recorded as the start time of the current hammer blow. When the standard penetration test hammer strikes, it is in free fall. There is a certain time interval before the worker pulls it up again. Therefore, the minimum static duration is set to 0.1s to 0.5s, which is sufficient to determine that the hammer has fallen to a standstill and that the time interval is less than the time interval between the worker's next pull-up operation.

[0081] S6. Using the original hammer height sequence from step S1 The vertical pixel coordinates of the bottom center point of the standard penetration test hammer bounding box at the end of the last hammer blow are obtained as the coordinates of the starting position of the current hammer blow. And obtain the end time of this hammer strike. The vertical pixel coordinates of the bottom center point of the corresponding standard penetration test hammer bounding box are used as the coordinates of the end position of this hammer strike. The difference between the two values ​​is calculated as the pixel displacement corresponding to this hammer strike:

[0082]

[0083] At the same time, the end time of this hammering will be... Update to the start time of the next hammer strike. This serves as the starting point for calculating the displacement of the next hammer blow.

[0084] S7. Scale calibration and depth conversion; uniformly select 10 frames containing complete reference rods from the video, and use existing HSV color segmentation and extraction algorithms to calculate the total pixel height of the reference rods in each frame. Frames that deviate from the mean by more than two standard deviations are removed, and the average value of the remaining frames is taken. Assuming the final pixel height; let the total physical height of the reference rod be... If the unit is cm, then the pixel-physical conversion factor is:

[0085] The unit is cm / pixel;

[0086] S8. Calculate the actual penetration depth for each hammer blow. for:

[0087] ;

[0088] System cumulative When the cumulative depth reaches 10 cm, 20 cm, and 30 cm respectively, the corresponding number of hammer blows is recorded to generate standard penetration test result data.

[0089] This invention relates to hammer impact event recognition and counting based on a state machine. To suppress target detection noise and environmental interference, the original height sequence is first smoothed using Gaussian filtering. Then, a motion logic-based state machine is constructed to determine the hammer's motion state at each moment (including "stationary," "falling," and "rising"). State determination employs a sliding window difference strategy: the change in Y-coordinate between the current frame and the previous 15 frames is calculated. If this change exceeds a preset threshold and meets the minimum motion speed requirement, it is considered a valid fall. After the fall process ends, if the position change is less than the stationary threshold for several consecutive frames (e.g., 3 frames), the hammer is considered to have stabilized, triggering a valid hammer impact count. Each time the hammer impact count is updated, the system automatically records the hammer position corresponding to the current and previous hammer impact moments, calculating the pixel displacement difference between the two as the basis data for subsequent depth conversion.

[0090] In this embodiment, the height reference rod is designed with equal-length red and green color blocks, each corresponding to a physical length of 60 centimeters. The overall structure is clear and has high contrast, facilitating image segmentation and recognition. During video processing, the system uniformly extracts 10 frames containing the complete reference rod over time, crops the reference rod region, and uses a color space segmentation algorithm (such as HSV thresholding) to accurately extract the pixel heights of the red and green segments in each frame. By removing outliers from the 10 frames (e.g., removing frames deviating from the mean by more than two standard deviations) and averaging the results, a robust average pixel height of the reference rod is obtained. Combined with its known total physical length, a conversion factor between pixels and centimeters is calculated. Finally, multiplying the pixel displacement corresponding to each hammer blow by this factor yields the actual penetration depth increment. Based on this, the number of hammer blows within each 10-centimeter depth segment is automatically accumulated, generating test result data that meets engineering specifications.

[0091] This invention achieves automation and traceability of standard penetration test data measurement through a four-step closed loop of "visual perception, state recognition, scale calibration, and physical conversion," without requiring contact sensors or changing existing drilling processes.

[0092] The above description is merely one embodiment of the present invention, and while it is detailed and specific, it should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A standard penetration test measurement method based on machine vision, characterized in that, The steps of the measurement method are as follows: S1. Install a height reference rod vertically next to the drill pipe of the standard penetration test. The height reference rod and the standard penetration hammer are on the same plane. Set up a high-resolution monocular camera 2.5 to 3.5 meters away from the test device to ensure that the hammer and the reference rod are in the center of the frame and unobstructed throughout the entire movement. Record the entire process of the standard penetration test with the camera. S2. Perform target detection frame-by-frame on all videos of the standard penetration test process to obtain the vertical pixel coordinates of the bottom center point of the SPT hammer bounding box at each moment during the standard penetration test. , The image frame number constitutes a sequence of the original hammer height at different times during the standard penetration test. ; S3. The original height sequence of the hammer in step S2. Each value in Gaussian smoothing is performed to eliminate high-frequency noise, resulting in a smoothed hammer height sequence. And the vertical pixel coordinates of the bottom center point of each standard penetration test hammer bounding box. The processing procedure is as follows: ； in, The vertical pixel value of the bottom center point of each SPT hammer bounding box after Gaussian smoothing; yes The original height of the hammer at time t; 'i' is a summation index representing an offset of the point coordinates around time t, with a value ranging from -N to N. The data from the preceding and following N points are used for weighted summation; Standard deviation Gaussian kernel; The kernel half-width is set to 30. S4. The motion trend is calculated using the sliding window difference method; the time window length is defined as a value 'a', which is between one-quarter and one-half of the shooting frame rate, and the smoothed hammer height difference between the current frame and the frame a before is calculated. : ； in, It is the height difference between the smoothed hammer height in frame t and the smoothed hammer height in frame a. It is the hammer height value in the t-th frame of the smooth sequence; In smooth sequences Frame hammer height value; according to The sign and amplitude of the hammer are combined with the current state of the hammer to dynamically determine the hammer's motion state, which includes falling, stationary, and rising. S5. When the state changes from falling to still and the minimum stillness duration is met, a valid hammer strike is determined to have occurred, and the current frame index is recorded as the end time of this hammer strike. The time when the last hammer blow ended is recorded as the start time of the current hammer blow. ; S6. Using the original hammer height sequence from step S1 The vertical pixel coordinates of the bottom center point of the standard penetration test hammer bounding box at the end of the last hammer blow are obtained as the coordinates of the starting position of the current hammer blow. And obtain the end time of this hammer strike. The vertical pixel coordinates of the bottom center point of the corresponding standard penetration test hammer bounding box are used as the coordinates of the end position of this hammer strike. The difference between the two values ​​is calculated as the pixel displacement corresponding to this hammer strike: ； At the same time, the end time of this hammering will be... Update to the start time of the next hammer strike. This serves as the starting point for calculating the displacement of the next hammer blow; S7. Scale calibration and depth conversion; uniformly select 10 frames from the video that contain complete reference rods, and calculate the total pixel height of the reference rods in each frame. Where i = 1, 2, ... 10, and outlier frames that deviate from the mean by more than 2 standard deviations are removed, and the average of the remaining frames is taken. Assuming the final pixel height; let the total physical height of the reference rod be... If the unit is cm, then the pixel-physical conversion factor is: The unit is cm / pixel; S8. Calculate the actual penetration depth for each hammer blow. for: ； System cumulative When the cumulative depth reaches 10 cm, 20 cm, and 30 cm respectively, the corresponding number of hammer blows is recorded to generate standard penetration test result data.

2. The standard penetration test measurement method based on machine vision according to claim 1, characterized in that: In step S2, a deep learning-based target detection model is used to simultaneously identify the standard penetration test hammer body and the dedicated height reference rod, obtain the coordinates of the lower left corner of the standard penetration test hammer bounding box, extract its vertical Y-axis position, and form a hammer height sequence that changes over time; the target detection model is trained using the YOLO model or the DETR model as the framework model.

3. The standard penetration test measurement method based on machine vision according to claim 1, characterized in that: In step S1, the camera records the entire experiment video at a frame rate of 30 frames per second; in step S4, the time window length is defined as 15 frames, and the height difference between the current frame and 15 frames ago is calculated. : 。 4. A standard penetration test measurement method based on machine vision according to claim 1 or 2, characterized in that, The height reference pole in step S1 consists of two equal-length red and green poles, each 60 cm high, for a total height of 120 cm.

5. A standard penetration test measurement method based on machine vision according to claim 1 or 2, characterized in that, The rules for determining the hammer's state in step S4 are as follows: Set the threshold based on the height reference pole pixel height obtained in step S7. and ; When in a "stationary" state: like The hammer body is determined to begin rising. like and The hammer is determined to have started falling. When in an "ascending" state: like This indicates that the upward trend has weakened or stopped, and the market has returned to a static state. If both conditions are met and If so, it is determined that the hammer body has changed from rising to falling; When in a "falling" state: like It is assumed that the fall is about to stop, so a stop timer is started; If this condition is satisfied continuously at least If a frame is reached, it is determined that the hammer has stabilized and stopped, and a valid hammer strike record is triggered. like This indicates that the hammer bounces back up immediately after falling, and the state switches to rising.

6. A standard penetration test measurement method based on machine vision according to claim 1 or 2, characterized in that: The minimum static duration in step S5 determines the time interval between the hammer falling and coming to a stop, and is less than the time interval between the worker's pull-up operation. It is set to 0.1s to 0.5s.

7. A standard penetration test measurement method based on machine vision according to claim 1 or 2, characterized in that: Step S7 uses the existing HSV color segmentation and extraction algorithm to calculate the total pixel height of the reference pole in each frame.

8. The standard penetration test measurement method based on machine vision according to claim 2, characterized in that: The target detection model adopts a target detection model based on the YOLOv8-small architecture. Its training process is as follows: First, image and video data containing the standard penetration test hammer, height reference rod and its working environment are collected at multiple standard penetration test sites, covering different lighting conditions, background complexity and hammer movement state; and the two types of targets in the images are manually labeled with annotation tools, namely "standard penetration hammer" and "height reference rod", to form an annotated training dataset. Subsequently, a transfer learning strategy was adopted, using the YOLOv8-small model pre-trained on a general image dataset as the initial weights, to fine-tune the training on the aforementioned special dataset. Furthermore, data augmentation methods are introduced during the training process to improve the model's adaptability to real-world working conditions; The data augmentation methods include image scaling, color perturbation, and geometric transformation; The model training is based on the detection accuracy on the validation set as the convergence criterion. Training is terminated when the performance tends to stabilize, thus obtaining a deep learning object detection model.

Images