A concrete pump truck pipe blockage cause behavior detection method and system based on posture key point guidance

Abstract

The present application relates to a kind of concrete pump truck pipe block inducement behavior detection method and system based on posture key point guide, construction site image is collected by camera, first, the target detection algorithm of state-posture auxiliary detection head is combined, the three-state detection of concrete pump truck is adjusted jib, pumping concrete, non-working state, and posture key point position information is simultaneously acquired;According to the spatial relationship of the last three key points, the soft tube region of interest is adaptively generated, and in the region, the normal state of the soft tube, patch state and joint state are distinguished and identified using multi-target semantic segmentation method.By introducing key point guide mechanism and mutual exclusion type semantic labeling rules, high-precision detection of concrete pump truck soft tube block inducement behavior is realized.This method has the advantages of high detection precision, low false detection rate and strong engineering adaptability, and is suitable for construction site concrete pump truck operation safety monitoring and intelligent early warning system.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and construction safety monitoring technology, and in particular to a method and system for detecting the causes of pipe blockage in concrete pump trucks based on attitude key point guidance. Background Technology

[0002] With the increasing height of buildings, pumped concrete has become the main method of concrete transportation. Pipe blockage is a key safety hazard that needs to be monitored during pumping. Severe blockage can lead to further pipe bursts, negatively impacting project progress, construction efficiency, and quality. Pump truck end hoses are typically 3 meters long, and splashed concrete aggregate from heights can endanger the lives and property of construction workers. Research shows that the risk of pipe blockage is directly related to pipe smoothness; pipe wear, sealing, and smoothness all indirectly affect the occurrence of accidents. According to the "Safety Regulations for the Use of Concrete Pumps and Pump Trucks in Construction Machinery and Equipment," "both the inside and outside of the hose should be visually inspected; if the reinforcing steel wires of the hose are visible, the hose should be scrapped." According to the "Safety Technical Regulations for the Use of Construction Machinery," "the placement piping and placement hose must not be extended." Therefore, this invention will use computer vision to detect illegal patching of hoses and the addition of pipes to the discharge port. Traditional manual inspections and video surveillance are limited by human resources and real-time response capabilities, resulting in problems such as delayed response and limited monitoring range. Furthermore, the personnel to be inspected may make temporary rectifications after receiving inspection notices, which cannot objectively reflect the safety of construction throughout the entire process.

[0003] Secondly, during construction, before and after the concrete pump truck has reached a stable pumping stage, workers often cover the discharge port with temporary plastic bags to prevent incomplete cleaning of the pump truck's pipelines. This can lead to concrete aggregate falling from the discharge port during boom adjustments or dripping onto the road during operation, potentially affecting road safety. While this is compliant with regulations, it can easily be misinterpreted as a hose patch or additional connection if the construction status of the concrete pump truck is not clearly identified.

[0004] Meanwhile, due to the complex environment of construction sites, the end hoses are often obstructed by passersby, construction workers, scaffolding and netting, building materials, steel bars, and other structures. In addition, the drone's camera position is not always fixed during inspections, resulting in diverse, sporadic, and temporary obstructions to the end hoses. Foreground obstructions disrupt the continuity of the hose's shape and may be mistaken for hose patches or additional connections. If these obstructions are not addressed, the engineering practicality of the invention will be greatly affected.

[0005] Therefore, there is an urgent need for a two-stage intelligent identification method for small targets, which can accurately identify whether a concrete pump truck is in pumping mode, and can also perform accurate, real-time, and comprehensive detection of small targets such as end hose patches and connecting pipes while in pumping mode, while eliminating the negative impact of foreground obstructions. This would improve the accuracy of identifying the causes of concrete pump truck blockage and the stability of the system, ensuring the safe operation of the construction site. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a method and system for detecting the causes of pipe blockage in concrete pump trucks based on attitude key point guidance. It only targets the pumping stage of concrete pump trucks and eliminates the influence of foreground objects blocking the flow. It designs a complete detection scheme for the causes of pipe blockage in hoses caused by patching damaged hoses and unauthorized addition of pipes. It has the advantages of high detection accuracy, low false detection rate, and strong engineering adaptability. It is suitable for the safety monitoring and intelligent early warning system of concrete pump truck operation at construction sites.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a method for detecting the causes of pipe blockage in concrete pump trucks based on attitude key point guidance, comprising the following steps:

[0008] Step S1: Collect aerial video of concrete pump trucks at the construction site using drones, cut it into several small videos, extract frames to construct a concrete pump truck action recognition image dataset, and complete the annotation and segmentation.

[0009] Step S2: YOLOv12n-Pose is used as the target detector. The backbone structure is optimized by introducing PEC (Perception enhancement convolution) and ESC2f (Enhanced Scope C2f) modules to improve feature fusion capability and build an improved detection network.

[0010] Step S3: Based on the improved detection network described in Step S2, a keypoint geometry encoder and a keypoint guided state classifier are further introduced to construct a state-pose assisted StateWithPose detection head, which is used to improve the model's accuracy in recognizing the working state of the concrete pump truck; the loss function is improved by weighting the key points of the end hose to focus on small targets on the end hose; the position and state of the concrete pump truck are output, namely, the boom adjustment, concrete pumping, and non-working states, and 9 posture key points of the concrete pump truck are identified;

[0011] Step S4: For the video frames identified as "pumping concrete" in Step S3, take the spatial location information of key points 7, 8, and 9 obtained in Step S3, and adaptively generate a region of interest covering the end hose; construct a multi-object segmentation dataset for anomaly detection of the end hose of the concrete pump truck, and complete pixel-level annotation of the hose, hose patch, and pipe connection area.

[0012] Step S5: Use SeaFormer for multi-target semantic segmentation to identify hoses, hose patches, and pipe-added areas in the dataset from Step S4. Check the consistency of the length of the connecting lines of key points 7, 8, and 9 before and after the images in the same short video. Finally, output the judgment result of the concrete pump truck blockage behavior caused by the presence or absence of patches and pipe-added areas.

[0013] In a preferred embodiment: In step S1, the image dataset includes 6144 images of concrete pump truck behavior with a resolution of 1920×1080 pixels, extracted from short videos taken by drones. Each short video contains at least 3 frames, and only one of the following states is taken: adjusting the boom, pumping concrete, or not working. The annotation format is YOLO format, and the dataset is divided into training set, validation set, and test set in a 5:1:1 ratio.

[0014] In a preferred embodiment: In step S2, the perceptual enhancement convolution PEC precisely cuts the input feature map into four small blocks with halved spatial dimensions and splices them together in the channel dimension to expand the receptive field and efficiently integrate contextual information; ESC2f, based on the C2f module, expands the feature map channel capacity, introduces a compression layer for intelligent channel filtering, and stacks multi-scale convolution kernels to fuse local and global features.

[0015] In a preferred embodiment: In step 3, the StateWithPose detection head is based on the YOLOv12n-Pose original detection head, and incorporates the arm length and joint angle geometric features of key points 1 to 7 into the judgment of the three states of the concrete pump truck, and acts on an independent learnable state classification branch.

[0016] In a preferred embodiment: In step S4, key point 7 represents the connection between the end hose and the boom, key point 9 represents the outlet of the end hose, and key point 8 represents the middle of the end hose; the spatial position information of key points 7, 8, and 9 refers to the left-right and up-down extreme values ​​of the three points; the adaptive method is to leave a margin of 0.03 times the vertical distance for the upper and lower extreme values, and leave a margin of 0.08 times the horizontal distance for the left and right extreme values, crop the region of interest, maintain the aspect ratio, resize the long side to 512 pixels, and fill the short side with black to 512 pixels, finally forming an end hose image of size 512×512 pixels.

[0017] In a preferred embodiment: In step S4, similar images are removed from the dataset and random permutation of RGB channels, raindrop noise is added, and random grayscale image data enhancement methods are used; the images are labeled as hose, patch, and pipe connection areas, wherein the overlapping parts of hose and patch, and hose and pipe connection are only labeled as patch and pipe connection; finally, 2492 sets of multi-target semantic segmentation images and mask labels for the causes of concrete pump truck pipe blockage are obtained.

[0018] In a preferred embodiment: In step S5, different frames of the same short video are set, and the consistency of the connection length of key points is checked. If the difference in length exceeds the threshold (30% of the maximum value of the connection), it is determined that there is a foreground object occlusion in the short video. There is a possibility of false detection or missed detection of the patching and splicing of the hose. In this case, it should be temporarily determined that there is no patching or splicing.

[0019] In a preferred embodiment: the method is ultimately used for detecting the causes of pipe blockage in concrete pump trucks at construction sites.

[0020] In a preferred embodiment: the short video is specifically a video with a duration of 3 seconds.

[0021] This invention also provides a concrete pump truck pipe blockage cause detection system based on attitude key point guidance, including a processor, a memory, and a bus. The memory stores machine-readable instructions executed by the processor. When the system is running, the processor communicates with the memory via the bus, and when the machine-readable instructions are executed by the processor, it is as described in the concrete pump truck pipe blockage cause detection method based on attitude key point guidance.

[0022] Compared with existing technologies, this invention has the following advantages: It fills the gap in monitoring construction safety of concrete pump trucks from a computer vision perspective. By using the CPTA-YOLOv12 concrete pump truck motion recognition detector and the Seaformer end hose anomaly detector, it constructs an intelligent monitoring method for the causes of pipe blockage in concrete pump trucks suitable for construction sites. This method maintains high recognition accuracy even in complex backgrounds and occluded environments. It boasts advantages such as high detection accuracy, lightweight and deployable model, and adaptability to diverse scenarios, making it suitable for higher-level supervision and construction site safety early warning systems for concrete pumping operations. Attached Figure Description

[0023] Figure 1 This is a flowchart of the overall system of the method of the present invention, which shows the key processing steps such as image acquisition, target and key point detection, result fusion and alarm.

[0024] Figure 2 This is the annotation method for the concrete pump truck action recognition dataset of this invention.

[0025] Figure 3 This invention describes the annotation method for a multi-object segmentation dataset used in the detection of abnormalities in the end hose of a concrete pump truck.

[0026] Figure 4 This is a structural diagram of the improved PEC module and ESC2f module used in this invention, which is used to improve the detection capability of small targets in the backbone network while controlling the number of parameters.

[0027] Figure 5 This is a diagram of the StateWithPose detection head structure proposed in this invention, which integrates a key point geometric encoder and a state classifier for optimizing the state classification branch.

[0028] Figure 6 This is a diagram of the improved YOLOv12n network structure of the present invention, showing the overall connection method of the backbone network, neck network and detection head module. Detailed Implementation

[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0030] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0031] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations according to this application; as used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise; furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0032] A method for detecting the causes of pipe blockage in concrete pump trucks based on attitude key point guidance, referenced Figure 1-6 This includes the following steps:

[0033] Step S1: Collect real-time video of concrete pump truck construction site by drone inspection and aerial photography, crop it into 3-second videos containing only a single action, extract frames from the 3-second videos to construct an image dataset of concrete pump truck actions, save the data in YOLO format after manual annotation, and finally form a concrete pump truck action recognition dataset of 6144 groups of 1920×1080 pixels, which is divided into training set, validation set and test set in a ratio of 5:1:1.

[0034] Step S2: Build the YOLOv12n-Pose object detector framework, introduce the Perceptual Enhancement Convolutional Module (PEC) to replace the first original convolutional module in the backbone, and use it to stitch together the channel dimensions of the image to expand the receptive field of the convolution by four times and efficiently integrate contextual information to improve the detection capability of small objects.

[0035] This invention uses CPTA-YOLOv12-Pose as the basic detection network structure, which mainly consists of three parts: an improved backbone network, a neck network, and an improved detection head. (See attached diagram.) Figure 5 As shown, the Backbone section introduces the Perceptual Augmentation Convolutional Module (PEC) and the Perceptual Augmentation Convolutional Module (ESC2f) to replace the first original convolutional module and the C3k2 module of the backbone, respectively (see attached diagram). Figure 4 The original YOLOv12n structure is maintained in the Neck stage; in the Head stage, a keypoint geometric encoder and a state classifier are introduced to construct a state-pose auxiliary head (see appendix). Figure 5 ), integrated into the StateWithPose-Detect structure (see appendix) Figure 6 In this study, the classification and regression branches are optimized to improve the detection capability of small targets and the structural stability.

[0036] The CPTA-YOLOv12-Pose model was trained using a YOLO format dataset, which contains 6144 images of a concrete pump truck in operation (attached). Figure 2 The ratio of the three working states (adjusting boom: pumping concrete: other working states) is 2130:2538:1476 (the geometric features of the skeletal points in the three states are a prerequisite for dividing the dataset: the skeletal point connections in adjusting the boom resemble broken lines, the skeletal point connections in pumping concrete resemble Bézier curves, and the skeletal point connections in other working states are stacked and move as a whole). The ratio of training set:test set:validation set is approximately 5:1:1. During training, the Adam optimizer is used with the ReLU activation function. The loss function is a multi-task weighted loss, including bounding box loss, pose keypoint loss, keypoint visibility loss, classification loss, distribution focus loss, and the newly added state classification loss and skeletal point consistency loss.

[0037] The specific steps are as follows:

[0038] Assume the input feature map is (H and W are the length and width of the feature map, and C is the number of channels in the feature map), and it is uniformly divided into 4 sub-blocks (X1, X2, X3, X4) in the spatial dimension (with a stride of 2). Each sub-block , as in equation (1).

[0039] (1)

[0040] The four sub-blocks are concatenated along the channel dimension, and the resulting feature map is: , as in equation (2).

[0041] (2)

[0042] Then, a 1×1 convolutional compression layer is used to reduce the channel dimension. After dimensionality reduction... To reduce computational load and retain key information, as in equation (3):

[0043] (3)

[0044] Finally, a Bottleneck module is used, which involves two 3×3 convolutions, and the input feature map is then concatenated with the residual of the feature map after the second convolution to further extract features, resulting in a perceptually enhanced feature map. As in equation (4):

[0045] (4)

[0046] ESC2f is introduced to replace the C3k2 module in the backbone, achieving multi-scale feature fusion to enhance the representation of small object features, while keeping parameter increases within a limited range. ESC2f introduces a three-branch multi-scale convolution on top of the original C2f, as follows:

[0047] The input feature map is ( (The number of input channels) is used to increase the channel capacity through a 1×1 convolution to obtain the feature map. ,in ( (where e is the number of output channels) and e is the expansion factor (default 0.5), as shown in equation (5):

[0048] (5)

[0049] Will The input consists of three branches. Each branch is first reduced to its dimensionality using a 1×1 convolution. aisle.

[0050] Branch 1 is the direct path, which, after dimensionality reduction, does not undergo any further convolutions to obtain the feature map. As shown in equation (6); branch 2 is a local feature path, which, after dimensionality reduction, undergoes two 3×3 convolutions to obtain the feature map. As shown in equation (7); branch 2 is the global feature path, which, after dimensionality reduction, undergoes a 5×5 convolution to obtain the feature map. As in equation (8):

[0051] (6)

[0052] (7)

[0053] (8)

[0054] The feature maps of the three branches are concatenated along the channel dimension to obtain Then, dimensionality reduction is achieved through 1×1 convolution. This achieves global feature fusion of shallow texture and deep semantic information, resulting in an enhanced range feature map. As in equation (9):

[0055] (9)

[0056] like (If the default condition is true), then residual connection is used, and the final output is .

[0057] Step S3: Based on the original YOLOv12n-Pose detection head structure, a state-pose assisted detection head is constructed by integrating a keypoint geometry encoder and a keypoint guided state classifier. This head is used to optimize the state classification branch, enabling the model to learn the relative geometric information of the skeletal points of the concrete pump truck excluding the hose. This further supplements and adjusts the image features of the concrete pump truck under the three states of boom adjustment, concrete pumping, and other states, thereby improving the accuracy of concrete pump truck action recognition.

[0058] The keypoint geometric encoder calculates the relative distance d between keypoints. ij and angle θ ij For example, equation (10) and equation (11):

[0059] (10)

[0060] (11)

[0061] The area and angle features are extracted from the key point triangles, and then the geometric relationships are encoded into feature vectors using a multilayer perceptron. (Geometric feature dimension) (Default value is 64). Assume the input tensor is... (B is the batch size, N is the number of anchor points, 7 is the total number of key points excluding the end hose, and 3 is the x, y coordinates and visibility of each key point), the key point geometric encoder expression is as shown in equation (12):

[0062] (12)

[0063] The state classifier can be divided into three layers. The first layer receives geometric features. and classification features obtained from the projection of detection features Perform feature-level fusion. After average pooling, the dimension becomes splicing and The fused features are obtained by fusing features through a multilayer perceptron (MLP). The specific calculation process for equation (13) is as follows:

[0064] (13)

[0065] The second layer is a preliminary state judgment based on fused features, that is... Unnormalized scores mapped to three state categories using a single-layer perceptron. As in equation (14):

[0066] (14)

[0067] The third layer uses a multi-layer perceptron to fuse the state prediction results from the three StateWithPose detection layers. (recorded as) , , That is, integrating the feature outputs from different receptive fields to finally predict the score, as shown in equation (15):

[0068] (15)

[0069] Step S4: Improve the loss function. The original loss function of Yolov12n-Pose is as shown in equation (16):

[0070] (16)

[0071] in It is the regression loss of the bounding box. It is the regression loss of the attitude key points. It is the loss of visibility at key points. It is classification loss. It is the distribution focus loss used to optimize bounding box regression.

[0072] The improved loss function is Internally, the weights of skeleton points 6, 7, and 8 are doubled, and a state classification loss is added. 2. Skeletal point geometric consistency loss As shown in equation (17):

[0073] (17)

[0074] Among them, the default weights of each loss are... , , , , , , The values ​​are 7.5, 10.0, 1.0, 0.5, 1.5, 0.5, and 0.15, respectively. The highest value for the keypoint loss represents the emphasis on end-cap hose detection, followed by the bounding box loss. The two newly added losses have relatively small weights, representing auxiliary tasks.

[0075] New state classification loss The loss function of the newly added StateWithPose detection head takes the geometric features of key points 1-7 and the detection classification features as input, and outputs three classification states to achieve the effect of geometric feature path-assisted classification verification and improve robustness, as shown in equation (18):

[0076] (18)

[0077] in, Labels for the actual state. Predict logits for the state.

[0078] The newly added geometric consistency loss of bone points Concrete pump truck operation samples in the same state are encouraged to have similar geometric features, as shown in equation (19):

[0079] (19)

[0080] Where P is the set of positive sample pairs (same state). For normalized geometric features, (The default value is 0.07) is the temperature parameter.

[0081] Step S5: For the video frame identified as "pumping concrete" in Step 3, take the left, right, top, and bottom extreme values ​​of bone points 7, 8, and 9, and record them as follows: ,exist Above and Take them separately The distance, in Left and Take again from the right The distance is calculated, and the video frames are cropped according to their regions of interest. After cropping, the image aspect ratio is maintained. The longer side is enlarged to 512 pixels, and the shorter side is padded with black pixels on both sides to reach 512 pixels, thus constructing the end-tight hose dataset image. Multi-object semantic segmentation uses a mutually exclusive semantic annotation rule, ensuring that each pixel corresponds to only one state semantic category. After some data augmentation, 2492 sets of end-tight hose anomaly detection dataset images are obtained, with the ratio of training set, validation set, and test set approximately 8:1:1.

[0082] Step S6: Use SeaFormer for multi-object semantic segmentation to identify hoses, hose patches, and pipe-added areas in the dataset S4. For images within the same 3-second video, check the consistency of the lengths of the lines connecting skeletal points 7, 8, and 9 before and after the images. If the difference between the maximum and minimum lengths is greater than 20% of the maximum, it is determined that there is foreground occlusion in that 3-second video, and no warning information is output for that 3-second video. Finally, combining the multi-object semantic segmentation detection results and the consistency check results, output the determination result of whether there are patches or pipe-added areas causing the concrete pump truck blockage.

[0083] This invention employs the Seaformer model for multi-object segmentation, used to identify patches and additional connectors in the region of interest (ROI) of the end-tight hose. Seaformer utilizes a lightweight compression-enhanced axial attention mechanism, reducing computational overhead through axial compression to capture the global hose structure, while simultaneously using local convolutional kernels to identify subtle textures of patches and additional connectors, resulting in significant efficiency advantages and suitability for deployment on edge devices. The model is trained on a self-built end-tight hose anomaly detection dataset (see attached dataset). Figure 3 The dataset includes 1236, 556, and 700 images and masks of normal hoses, hoses with patches, and hoses with added connectors, respectively. The training set:validation set:test set ratio is 8:1:1. The average Dice coefficient (mDice) and classification accuracy (mAcc) of each category of pixels are obtained. Finally, by combining multi-target semantic segmentation detection and skeletal point connection consistency checks, the results are output to determine the cause of concrete pump truck pipe blockage behavior with or without patches and added connectors.

Claims

1. A method for detecting the causes of pipe blockage in concrete pump trucks based on attitude key point guidance, characterized in that, Includes the following steps: Step S1: Collect aerial video of concrete pump trucks at the construction site using drones, cut it into several small videos, extract frames to construct a concrete pump truck action recognition image dataset, and complete the annotation and segmentation. Step S2: Using YOLOv12n-Pose as the target detector, the backbone structure is optimized, and perceptual enhancement convolution PEC and enhanced range C2f modules are introduced to improve feature fusion capabilities, thus constructing an improved detection network; Step S3: Based on the improved detection network described in Step S2, a key point geometric encoder and a state classifier are further introduced to construct a state-pose assisted StateWithPose detection head, which is used to improve the model's recognition accuracy of the concrete pump truck's working state; the loss function is improved by weighting the key points of the end hose to focus on small targets on the end hose; the position and state of the concrete pump truck are output, namely, adjusting the boom, pumping concrete, and non-working states, and 9 posture key points of the concrete pump truck are identified. Step S4: For the video frames identified as "pumping concrete" in Step S3, take the spatial location information of key points 7, 8, and 9 obtained in Step S3, and adaptively generate a region of interest covering the end hose; construct a multi-object segmentation dataset for anomaly detection of the end hose of the concrete pump truck, and complete pixel-level annotation of the hose, hose patch, and pipe connection area. Step S5: Use SeaFormer for multi-target semantic segmentation to identify hoses, hose patches, and pipe-added areas in the dataset from Step S4. Check the consistency of the length of the connecting lines of key points 7, 8, and 9 before and after the images in the same short video. Finally, output the judgment result of the concrete pump truck blockage behavior caused by the presence or absence of patches and pipe-added areas.

2. The method for detecting the causes of pipe blockage in concrete pump trucks based on attitude key point guidance according to claim 1, characterized in that: In step S1, the image dataset includes 6144 images of concrete pump truck behavior with a resolution of 1920×1080 pixels, extracted from short videos taken by drones. Each short video contains at least 3 frames, showing only one of the following states: adjusting the boom, pumping concrete, or not working. The images are labeled in YOLO format and divided into training, validation, and test sets in a 5:1:1 ratio.

3. The method for detecting the causes of pipe blockage in concrete pump trucks based on attitude key point guidance according to claim 1, characterized in that: In step S2, the Perceptual Enhancement Convolution (PEC) precisely cuts the input feature map into four spatially halved blocks and splices them in the channel dimension to expand the receptive field and efficiently integrate contextual information; the Enhanced Range (C2f) module expands the feature map channel capacity, introduces a compression layer for intelligent channel filtering, and stacks multi-scale convolutional kernels to fuse local and global features based on the C2f module.

4. The method for detecting the causes of pipe blockage in concrete pump trucks based on attitude key point guidance according to claim 1, characterized in that: In step 3, the StateWithPose detection head is based on the YOLOv12n-Pose original detection head, and incorporates the arm length and joint angle geometric features of key points 1 to 7 into the judgment of the three states of the concrete pump truck, and acts on an independent learnable state classification branch.

5. The method for detecting the causes of pipe blockage in concrete pump trucks based on attitude key point guidance according to claim 1, characterized in that: In step S4, key point 7 represents the connection between the end hose and the boom, key point 9 represents the outlet of the end hose, and key point 8 represents the middle of the end hose. The spatial position information of key points 7, 8, and 9 refers to the left-right and up-down extreme values ​​of the three points. The adaptive method is to leave a margin of 0.03 times the vertical distance for the upper and lower extreme values, and leave a margin of 0.08 times the horizontal distance for the left and right extreme values. The region of interest is cropped, the aspect ratio is maintained, the long side is resized to 512 pixels, and the short side is filled with black to 512 pixels, finally forming an end hose image of size 512×512 pixels.

6. The method for detecting the causes of pipe blockage in concrete pump trucks based on attitude key point guidance according to claim 1, characterized in that: In step S4, similar images are removed from the dataset, and random permutation of RGB channels, addition of raindrop noise, and random grayscale image data augmentation methods are used. The images are labeled with hose, patch, and pipe connection areas, and the overlapping parts of hose and patch, and hose and pipe connection are only labeled with patch and pipe connection. Finally, a total of 2492 sets of multi-target semantic segmentation images and mask labels for the causes of concrete pump truck pipe blockage are obtained.

7. The method for detecting the causes of pipe blockage in concrete pump trucks based on attitude key point guidance according to claim 1, characterized in that: In step S5, different frames of the same short video are set up to check the consistency of the connection length of key points. If the difference in length exceeds the threshold, it is determined that there is a foreground object occlusion in the short video. There is a possibility of false detection or missed detection of the patching and splicing of the hose. In this case, it should be temporarily determined that there is no patching or splicing.

8. A method for detecting the causes of pipe blockage in concrete pump trucks based on attitude key point guidance according to any one of claims 1 to 7, characterized in that: The method was ultimately used to detect the causes of pipe blockage in concrete pump trucks at construction sites.

9. A method for detecting the causes of pipe blockage in concrete pump trucks based on attitude key point guidance according to any one of claims 1 to 7, characterized in that: The short video is specifically a video with a duration of 3 seconds.

10. A concrete pump truck blockage-inducing behavior detection system based on attitude key point guidance, comprising a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executed by the processor; characterized in that, When the system is running, the processor and the memory communicate via a bus, and the machine-readable instructions are executed by the processor as described in any one of claims 1 to 7.

Images