Concrete construction water adding quality management system and method based on deep learning

Abstract

This invention discloses a deep learning-based quality management system and method for water addition in concrete construction, belonging to the field of concrete construction quality management technology. The method includes: extracting keyframes and decomposing spatial domain features from continuous video of the work site to obtain color space and edge gradient components; generating a set of on-site object distributions based on target contour morphology recognition; tracking the positional offset of objects between adjacent frames to obtain a sequence of motion trajectories for each target; projecting the trajectories onto a three-dimensional spatial coordinate system centered on the hopper opening to obtain a trajectory spatial distribution matrix; performing temporal analysis on the positional overlap between the trajectories and the hopper opening area to obtain an overlap time series; and cumulatively calculating the dwell time of each target in the hopper opening area, triggering a water addition warning when a threshold is exceeded and pushing it to the management terminal. This invention achieves automatic identification and early warning of illegal water addition by concrete pump trucks.

Description

Technical Field

[0001] This invention relates to the field of concrete construction quality management technology, specifically to a concrete construction water addition quality management system and method based on deep learning. Background Technology

[0002] During concrete pump truck operations, the unauthorized addition of water to the hopper alters the water-cement ratio of the concrete, leading to a decrease in structural strength and posing a quality hazard. Current on-site water addition monitoring relies primarily on manual inspections and fixed-location video surveillance. Manual inspections suffer from blind spots and response delays, making it difficult to continuously monitor multiple pump trucks simultaneously. Fixed surveillance footage only provides two-dimensional images, failing to accurately reconstruct the spatial relationship between personnel and water-adding equipment near the hopper opening. When personnel approach the hopper, it's difficult to distinguish between normal operations and water-adding preparations, resulting in a high false alarm rate. Some solutions utilize target detection algorithms to identify personnel and water-adding equipment, but based solely on single color or shape features, the differentiation between worker skin color areas and water-adding equipment color areas is unstable in construction environments with varying lighting and dust obstruction. The outline segmentation is coarse, easily mistaking clothing, equipment shadows, etc., as targets. Furthermore, existing solutions lack correlation analysis of the carrying relationship between personnel and water-adding equipment. Isolated target detection results cannot confirm whether the water-adding equipment is carried by a specific person towards the hopper opening, making it impossible to accurately identify the violator.

[0003] Furthermore, when determining whether water addition actually occurs, existing methods rely solely on the simple overlap between the target bounding box and a fixed area, without quantifying the continuous dwell time of the trajectory in three-dimensional space. This can easily generate false alarms for scenarios where personnel briefly pass through the hopper or merely approach an area without water being added. It also fails to provide precise data on the duration of dwell time, resulting in insufficient reliability of the warning. On-site management personnel need to verify the information frequently, and their workload has not been substantially reduced. Summary of the Invention

[0004] This invention provides a deep learning-based method for water addition quality management in concrete construction. It aims to solve the technical problems of existing supervision methods in complex construction environments, such as inaccurate identification of water addition behavior, inability to reliably associate the relationship between personnel and water addition equipment, and inability to accurately quantify the water addition retention process. By using joint feature decomposition of image color space and edge gradient, it achieves stable segmentation and pairing of personnel, water addition equipment, and pump truck equipment areas. Furthermore, by utilizing the temporal overlap analysis of three-dimensional spatial trajectory and hopper opening area, it enables automatic calculation and high-reliability early warning of the retention time of illegal water addition behavior.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a method for quality management of water addition in concrete construction based on deep learning, comprising: acquiring continuous video stream data of a concrete pump truck operation site, performing keyframe extraction processing to obtain an image sequence to be analyzed and performing spatial domain feature decomposition processing to extract the color space component and edge gradient component of each pixel in the image; performing contour morphology recognition processing on target objects in the image sequence to be analyzed based on the color space component and edge gradient component to generate a set of on-site object distributions including personnel posture features and object carrying features; and performing temporal trajectory tracking processing on the object position offset between adjacent frames in the image sequence to be analyzed based on the on-site object distribution set. The method involves obtaining motion trajectory sequences for each target object; performing spatial coordinate mapping on these motion trajectory sequences, projecting them onto a three-dimensional spatial coordinate system centered on the pump truck hopper opening, generating a trajectory spatial distribution matrix; based on this matrix, performing positional overlap time-series analysis on the motion trajectories of each target object and a preset hopper opening spatial region, generating a trajectory-region overlap time series; performing cumulative duration calculation on the trajectory-region overlap time series, generating statistical values ​​of the dwell time of each target object within the hopper opening spatial region; when the dwell time statistical value exceeds a preset duration alarm threshold, triggering a water addition behavior warning signal, and pushing the warning signal to the on-site management terminal. This method, through multi-source image feature fusion and three-dimensional spatial behavior analysis, achieves non-contact real-time monitoring and automatic evidence collection of illegal water addition behavior in concrete construction, with high detection accuracy and strong resistance to environmental interference.

[0006] As a preferred technical solution of the present invention, each frame of the image sequence to be analyzed is input into a pre-trained deep convolutional neural network, and a multi-channel feature map output by a convolutional layer in the deep convolutional neural network is extracted; the channel in the multi-channel feature map that responds to color distribution information is taken as the color space component, and the channel that responds to edge gradient information is taken as the edge gradient component.

[0007] As another preferred technical solution of the present invention, the color space components and edge gradient components are input into a region proposal network to generate a set of candidate target boxes; the set of candidate target boxes is processed using a classification and regression network to identify personnel targets and water supply equipment targets, and the corresponding bounding box coordinates and category confidence scores are output; a first descriptor representing the posture characteristics of personnel and a second descriptor representing the object-carrying characteristics of water supply equipment are extracted based on the bounding box coordinates; the identified personnel targets and water supply equipment targets are paired and associated according to their spatial proximity to generate the on-site object distribution set containing the relationship between personnel and objects.

[0008] Preferably, the temporal trajectory tracking processing includes: starting from the initial frame of the image sequence to be analyzed, extracting the centroid coordinates and contour dimensions of each target object in the on-site object distribution set to construct a target state vector set for the initial frame; for each target object in the current frame, calculating the centroid Euclidean distance and contour intersection-union ratio between it and all candidate target objects in the next frame to form an inter-frame matching cost matrix; performing Hungarian assignment processing on the inter-frame matching cost matrix to determine the optimal pairing relationship between the target object in the current frame and the candidate target object in the next frame; according to the optimal pairing relationship, appending the centroid coordinates of the successfully paired candidate target objects in the next frame to the end of the trajectory point sequence of the corresponding target object; when a target object cannot establish a pairing relationship with any candidate target object in the next frame, terminating the trajectory tracking of the target object, and outputting the accumulated trajectory point sequence as the motion trajectory sequence of the target object. This method can maintain trajectory continuity when the target undergoes brief occlusion or appearance changes, significantly improving the robustness of trajectory tracking.

[0009] Preferably, the spatial coordinate mapping process includes: extracting the pixel horizontal and vertical coordinates of each trajectory point in the motion trajectory sequence in the image plane coordinate system, and obtaining the timestamp information corresponding to the trajectory point; obtaining a pre-calibrated perspective projection transformation matrix between the image pixel coordinates and the three-dimensional spatial coordinates; inputting the pixel horizontal and vertical coordinates and the corresponding timestamp information of each trajectory point into the perspective projection transformation matrix for linear transformation processing to obtain the three-dimensional spatial coordinate values ​​of the trajectory point in the three-dimensional spatial coordinate system centered on the pump truck hopper opening; arranging all the three-dimensional spatial coordinate values ​​of the same target object into a three-dimensional trajectory vector according to the timestamp order, and combining the three-dimensional trajectory vectors of all target objects into the trajectory spatial distribution matrix. The perspective projection transformation restores the two-dimensional image trajectory to the three-dimensional space of the real construction scene, providing an accurate coordinate reference for subsequent spatial overlap analysis.

[0010] Furthermore, the position overlap time sequence analysis includes: extracting the three-dimensional spatial coordinate values ​​corresponding to each trajectory point of each target object from the trajectory spatial distribution matrix; comparing each three-dimensional spatial coordinate value with the boundary coordinate range of the hopper opening spatial region point by point to determine whether the trajectory point falls inside the hopper opening spatial region; for each target object, recording the time values ​​corresponding to the trajectory points falling inside the hopper opening spatial region in chronological order to form an initial time marker sequence for the target object; performing interval merging processing on the continuous time values ​​in the initial time marker sequence, merging time points with adjacent time differences less than the frame interval threshold into the same overlapping time interval; arranging the start and end times of all overlapping time intervals in chronological order to generate the trajectory-region overlap time sequence. This method can effectively distinguish between the worker's brief passage behavior and continuous water-adding preparation actions, reducing the probability of false alarms.

[0011] Furthermore, the generation of the retention time statistics includes: extracting the start and end times of each overlapping time interval from the trajectory-region overlapping time series; performing a difference operation by subtracting the start time from the end time for each overlapping time interval to obtain the single retention time corresponding to that overlapping time interval; summing the single retention times of all overlapping time intervals corresponding to the same target object to obtain the cumulative retention time of the target object in the hopper opening space area; and summing the cumulative retention times of all target objects in the trajectory-region overlapping time series to generate the retention time statistics. Using the cumulative retention time as the basis for early warning is more in line with the actual behavior characteristics of workers adding water in small amounts multiple times at the construction site, thus improving the rationality of the early warning.

[0012] Preferably, the process of triggering a water-addition behavior warning signal includes: comparing the statistical value of the dwell time with the duration alarm threshold to determine whether the statistical value of the dwell time is greater than the duration alarm threshold; when the statistical value of the dwell time is greater than the duration alarm threshold, determining the identity of the target object corresponding to the current time point and the trajectory spatial coordinates of the target object in the hopper opening space area; based on the trajectory spatial coordinates, extracting a video segment from the continuous video stream data between the moment the target object enters the hopper opening space area and the current moment as video evidence of the violation; packaging the video evidence of the violation, the identity of the target object, and the current timestamp into the warning signal, and pushing it to the on-site management terminal via a wireless communication link. This automatically generates warning information containing complete video evidence, facilitating rapid response and post-event traceability for management personnel.

[0013] This invention also provides a machine vision-based concrete construction water addition quality management system, including a processor and a memory. The memory is connected to the processor and is used to store programs, instructions, or code. The processor executes the programs, instructions, or code in the memory to implement the aforementioned deep learning-based concrete construction water addition quality management method. This system is flexibly deployable and can be directly connected to existing video surveillance networks to achieve 24 / 7 unattended intelligent monitoring and management of water addition activities.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By extracting the image sequence to be analyzed from a continuous video stream and decomposing the color space components and edge gradient components of each pixel, pixel-level classification is performed on the worker's skin color region, the water supply equipment color region, and the pump truck equipment region using the hue component. Simultaneously, an initial set of target candidate boxes is generated by adaptive binarization and connected component labeling of the edge gradient components. Then, the positional overlap rate between the worker's skin color region and the water supply equipment color region and the candidate box set is used as a constraint to filter out the personnel posture features and the features carried by the water supply equipment. The two are then paired and associated according to spatial proximity to form a set of on-site object distributions that includes the relationship between personnel and the objects they carry. This method complementarily integrates the semantic information of color segmentation and the geometric information of edge contours. It can reliably distinguish the exposed skin of the human body, the material color of the water supply equipment, and the structural edges of the pump truck even under changes in lighting and dust interference. Furthermore, based on spatial proximity pairing, it can clearly establish the carrying relationship between personnel and water supply equipment, avoiding misjudging unrelated personnel and water supply equipment as water supply subjects, thus reducing false alarms from the source of object detection.

[0015] By mapping the motion trajectory sequences of various target objects to spatial coordinates, and using a pre-calibrated perspective projection transformation matrix, the trajectory points are transformed from image plane coordinates to a three-dimensional spatial coordinate system centered on the hopper opening of the pump truck, generating a trajectory spatial distribution matrix. Then, the three-dimensional trajectory points of each target object are superimposed on the boundary of the hopper opening spatial area point by point. The time values ​​of falling into the area are recorded in chronological order, and consecutive moments are merged into overlapping time intervals. Finally, the duration of each overlapping interval is accumulated to obtain a statistical value of the dwell time. When the threshold is exceeded, video clips are extracted and packaged with the target object's identification information to send an alert. This method restores the motion trajectory in the two-dimensional image to the true three-dimensional space, eliminating distance misjudgments caused by perspective distortion, accurately distinguishing whether a person is passing by the side of the hopper opening or extending into the space above the hopper opening. Through continuous merging and accumulation of overlapping time intervals, an alarm is triggered only when a person carrying water-filling equipment continuously stays in the hopper opening area for a specified duration, filtering out false alarms caused by short-term passing or simple proximity, thus providing temporal and spatial quantitative support for evidence of illegal water-filling behavior. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of a deep learning-based method for managing the quality of water addition during concrete construction. Figure 2 This is a flowchart of the process for generating the on-site object distribution set; Figure 3 This is a flowchart of the target object motion trajectory tracking process; Figure 4 It is a schematic diagram of the target object's movement trajectory; Figure 5 It is a statistical curve of the duration of the target object's stay in the space area of ​​the hopper opening. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0019] See Figure 1 This invention provides a deep learning-based method for quality management of water addition during concrete construction, comprising: Continuous video stream data of the concrete pump truck operation site is acquired, key frame extraction is performed, the image sequence to be analyzed is obtained, and spatial domain feature decomposition is performed to extract the color space component and edge gradient component of each pixel in the image. Based on the color space components and edge gradient components, the target objects in the image sequence to be analyzed are subjected to contour morphology recognition processing to generate a set of on-site object distributions. Based on the set of on-site object distribution, temporal trajectory tracking processing is performed on the object position offset between adjacent frames in the image sequence to be analyzed to obtain the motion trajectory sequence of each target object. The motion trajectory sequence is processed by spatial coordinate mapping, and the motion trajectory sequence is projected onto a three-dimensional spatial coordinate system constructed with the pump truck hopper opening as the center to generate a trajectory spatial distribution matrix; Based on the trajectory spatial distribution matrix, the positional overlap time sequence analysis of the motion trajectory of each target object and the preset hopper opening spatial region is performed to generate the trajectory-region overlap time sequence; The trajectory-region overlap time series is processed by cumulative duration calculation to generate statistical values ​​of the residence time of each target object in the hopper opening space area; When the statistical value of the dwell time exceeds the preset duration alarm threshold, a water addition behavior warning signal is triggered and the warning signal is pushed to the on-site management terminal.

[0020] Example 1: In a specific implementation, the process of extracting the color space components and edge gradient components of each pixel in the image is achieved through the following steps.

[0021] Each frame of the image sequence to be analyzed is input into a pre-trained deep convolutional neural network (DCNN). The image is then forward-propagated to intermediate convolutional layers, and multi-channel feature maps output from these layers are extracted. Each multi-channel feature map contains several channels, and each channel represents a two-dimensional activation response map. In some embodiments, the DCNN uses a VGG-16 network structure, extracting the multi-channel feature map output from the fifth convolutional layer, where the channel dimension is 256. In other embodiments, the DCNN uses a ResNet-50 network structure, extracting the multi-channel feature map output from the convolutional layers after the fourth residual block, where the channel dimension is 1024. Optionally, the extraction of multi-channel feature maps can be combined with a feature pyramid strategy. The feature maps output from the third, fifth, and eighth convolutional layers of the DCNN are upsampled and their channels are reduced in dimension before being summed to generate a multi-channel feature map that integrates multi-scale information. Optionally, the extraction of multi-channel feature maps can also involve concatenating the high-resolution feature map output from the second convolutional layer in a deep convolutional neural network with the high-level semantic feature map output from the thirteenth convolutional layer along the channel dimension to obtain a multi-channel feature map that simultaneously preserves fine color details and abstract edge semantics.

[0022] When performing channel attribute partitioning on a multi-channel feature map, based on the degree of response bias of each channel to color distribution information and edge gradient information, all channels are divided into color space component channel groups and edge gradient component channel groups. A response attribute evaluation function based on a labeled mask is used for this partitioning. For the c-th channel in the multi-channel feature map, the color response bias... The formula is:

[0023] In the formula, This represents the set of coordinates of all pixels within the color distribution information annotation area. Label the color distribution mask in coordinates The value at that location, coordinates If the value is within the marked color area, use 1; otherwise, use 0. For the c-th channel in the multi-channel feature map, the coordinates are... The activation response value at that location, Represents the set of pixel coordinates for the entire image. This is a set small constant; simultaneously, the edge response bias is calculated using the same formula, replacing the color distribution information annotation area with the edge gradient annotation area. When the color response bias of the c-th channel... When the edge response bias of the c-th channel exceeds a preset color bias threshold, the c-th channel is assigned to the color space component. When the edge response bias of the c-th channel exceeds a preset edge bias threshold, the c-th channel is assigned to the edge gradient component. It can be understood that the color space component concentrates the channels that respond to color information such as the orange reflective vest of the construction worker, the blue water supply equipment, and the gray concrete, while the edge gradient component concentrates the channels that respond to edge gradient information such as the edge of the hopper opening, the cylindrical outline of the water supply equipment, and the outline of the worker's limbs. The feature vectors provided by the color space component and the edge gradient component at the same pixel location are independent and complementary.

[0024] In an image taken from a concrete pump truck operation site, a construction worker wearing an orange reflective vest stands to the right of a gray hopper opening, holding a blue polyethylene water dispensing device in his right hand. Gray, wet concrete is scattered around the hopper opening. The image is input into a deep convolutional neural network with a VGG-16 structure, and the 256-channel multi-channel feature map output from the fifth convolutional layer is extracted. After being partitioned by a response attribute evaluation function, 128 color space component channels and 128 edge gradient component channels are obtained. In the 23rd channel of the color space component, the average activation response value of pixels in the blue water dispensing device area is 0.86, while the average activation response value of pixels in the gray surface area of ​​the hopper opening is only 0.14, showing a significant difference. In the 89th vertical gradient response channel of the edge gradient component, the activation response value of pixels at the left and right boundaries of the water dispensing device reaches 0.79, while the activation response value of the homogeneous area inside the water dispensing device is less than 0.10. For data comparison, without distinguishing between color space components and edge gradient components, directly concatenating 256-channel multi-channel feature maps for feature representation resulted in a ratio of 1.3 between the inter-class and intra-class scatter matrix traces of the feature vector in the same area where a blue water-filling device meets a gray background. However, when using color space components alone to represent the water-filling device area, this ratio increased to 3.2. When using edge gradient components alone to represent the outline of the water-filling device, this ratio increased to 3.5. By acquiring color space components and edge gradient components separately, the distinguishability of feature descriptions for water-filling devices and personnel targets was maintained. Even when faced with interference from surrounding concrete water stains or non-uniform lighting, specific color response channels in the color space components stably captured the intrinsic colors of the object, while multi-directional gradient channels in the edge gradient components accurately depicted the geometric boundaries of the object, reducing the possibility of missed and false detections. It is understandable that in the subsequent contour shape recognition processing stage, the color space components and edge gradient components can be fed into the region proposal network using differential input channels or weighted fusion, so that the recall rate and position accuracy of the target box proposal are not degraded due to the changing on-site lighting and the gray cement dust in the hopper opening area.

[0025] Example 2: In specific implementation, refer to Figure 2 The target objects in the image sequence to be analyzed are subjected to contour morphology recognition processing based on the color space components and edge gradient components to generate a set of on-site object distributions. The implementation process is as follows.

[0026] Color space components and edge gradient components are input into a region proposal network, which consists of several convolutional layers and a multi-channel input adaptation layer. The multi-channel input adaptation layer concatenates the color space components and edge gradient components in the channel dimension to form a joint feature tensor. Subsequently, the region proposal network uses a sliding window approach to generate anchor boxes of various scales and aspect ratios at each spatial location of the joint feature tensor and predicts the foreground and background scores and coordinate offsets for each anchor box. Anchor boxes whose overlapping area with the image boundary exceeds a preset area threshold are removed, and after non-maximum suppression processing, a set of candidate target boxes is obtained. In a specific image acquisition example, a frame of an image from a concrete pump truck operation site contains a construction worker wearing an orange reflective vest and carrying a blue water dispenser in his right hand. The orange channel in the color space component shows a high activation response in the water dispenser area, while the vertical edge channel in the edge gradient component clearly delineates the cylindrical boundaries on the left and right sides of the water dispenser. After receiving the color space component and the edge gradient component, the region proposal network densely generates multiple candidate bounding boxes near the outline of the construction worker's torso and limbs, as well as around the water dispenser. Some candidate bounding boxes tightly surround the construction worker's body, while others tightly surround the water dispenser area. Each candidate bounding box in the set is accompanied by candidate bounding box coordinates and a foreground confidence estimate.

[0027] In practical implementation, a classification and regression network is used to process the candidate bounding box set. This network includes a region-of-interest pooling layer, a fully connected layer, and parallel classification and regression output layers. The network extracts in-box features from each candidate bounding box, determines its category, and refines its bounding box. It identifies two object categories: personnel and water supply equipment. It outputs the bounding box coordinates and category confidence score for each personnel target, and the bounding box coordinates and category confidence score for each water supply equipment target. In the aforementioned image containing construction workers and water supply equipment, the network classifies candidate bounding boxes tightly surrounding the construction workers as personnel targets and outputs a personnel category confidence score of 0.94 and the corrected bounding box coordinates based on the regression offset. Similarly, it classifies candidate bounding boxes tightly surrounding the water supply equipment as water supply equipment targets and outputs a water supply equipment category confidence score of 0.91 and the corrected bounding box coordinates based on the regression offset. Based on the corrected personnel bounding box coordinates, a first descriptor representing the personnel's posture features is extracted. The first descriptor is obtained by applying a human key point heatmap regression model to the internal region of the personnel bounding box. The human key point heatmap regression model outputs the two-dimensional coordinates of the construction worker's shoulder joints, arm key points, and wrist joints. The angle between the upper arm and forearm and the overall deflection angle of the arm relative to the vertical direction are calculated from the two-dimensional coordinates of the joints, forming a fourteen-dimensional personnel posture description vector. Based on the corrected water filling equipment bounding box coordinates, a second descriptor representing the carrying features of the water filling equipment is extracted. The second descriptor is obtained by applying a symmetry axis detection and handle position positioning model to the internal region of the water filling equipment bounding box. The symmetry axis detection and handle position positioning model outputs the tilt angle of the water filling equipment's symmetry axis and the offset vector of the handle connection point relative to the bucket opening. The tilt angle and offset vector form a six-dimensional water filling equipment carrying feature vector.

[0028] In practice, all identified personnel targets and all water supply equipment targets are paired and associated according to their spatial proximity. The association strength evaluation value between each pair of personnel and water supply equipment targets is calculated. in accordance with:

[0029] In the formula, This represents the Euclidean distance between the center point of the bounding box of the personnel target involved in the current pairing and the center point of the bounding box of the water supply equipment target. This is a distance normalization constant pre-calibrated based on the average physical scale of the working area around the hopper opening. This indicates the category confidence level of the personnel targets involved in the current pairing. This indicates the category confidence of the water supply equipment target involved in the current pairing. When the association strength evaluation value calculated for a pair of personnel targets and water supply equipment targets exceeds the preset association threshold of 0.6, it is determined that the personnel target carries the water supply equipment target, and the personnel target and the water supply equipment target are associated as a set of personnel-carried object relationships. In a complex scene where two construction workers and two water-filling devices appear simultaneously in a single frame, the aforementioned processing was performed using a joint input method of color space components and edge gradient components. The first construction worker was correctly paired with the blue water-filling device next to him, and the second construction worker was correctly paired with the gray water-filling device next to him, resulting in two correct pairings without any false alarms. However, in a contrast processing method that masks the edge gradient components and retains only the color space components, due to the lack of geometric constraints on the water-filling device outlines, the edge responses of a blue water-filling device and a gray water-filling device in the vicinity of a construction worker wearing a blue shirt became confused. The candidate bounding boxes generated by the region proposal network contained deformed candidate boxes that mistakenly classified the edge of the blue water-filling device into the boundary of the construction worker's shirt. This resulted in the same construction worker being simultaneously associated with two water-filling devices after the association strength evaluation value was calculated, reducing the number of correct pairings to one and generating one false alarm pairing. The bounding box coordinates, category confidence, first descriptor, second descriptor, and the relationship of the objects carried by the person and the water-filling device targets that were successfully paired were combined into a field object distribution entry. All field object distribution entries were then summarized to generate a field object distribution set. It is understandable that the on-site object distribution set records the pairing and binding relationship between personnel targets and water supply equipment targets in each frame of the image in a structured manner, providing an object-level identifier for each personnel target and the associated water supply equipment target for subsequent inter-frame trajectory tracking.

[0030] Example 3: In specific implementation, refer to Figure 3 Based on the set of on-site object distributions, temporal trajectory tracking processing is performed on the object position offsets between adjacent frames in the image sequence to be analyzed, resulting in a motion trajectory sequence for each target object. The implementation process is as follows.

[0031] Starting from the initial frame of the image sequence to be analyzed, obtain the on-site object distribution set corresponding to the initial frame. The on-site object distribution set contains several on-site object distribution records, each corresponding to a target object. The target object can be a combination of paired and associated personnel and water supply equipment targets. Extract the centroid coordinates and contour dimensions of the target object from each on-site object distribution record. The centroid coordinates are determined by the average of the centroid coordinates of the personnel candidate frame and the water supply equipment candidate frame. The contour dimensions are the width and height of the bounding rectangle formed by merging the bounding rectangles of the personnel and water supply equipment candidate frames. Each target object's centroid coordinates is represented as a three-dimensional vector, its components including the image plane's horizontal and vertical coordinates and the initial timestamp, which is the acquisition time of the initial frame. The contour dimensions are represented as a two-dimensional vector, its components including the width and height of the composite bounding rectangle. A target object's centroid coordinate vector and contour dimension vector are combined to form a target state vector. Collect all target state vectors from the initial frame to construct the initial frame target state vector set.

[0032] For each target object in the current frame, calculate the centroid Euclidean distance and contour intersection-over-union ratio between this target object and all candidate target objects in the next frame. Let the x-th target object in the current frame be the x-th target object. The centroid of the target object is In the next frame The centroid of each candidate target object is The first in the current frame The target object and the first in the next frame Euclidean distance between the centroids of candidate target objects Defined as:

[0033] in, For the current frame, the first The target object and the first in the next frame The centroid Euclidean distance between candidate target objects Indicates the first in the current frame The x-coordinate of the centroid of the target object in the image plane Indicates the first in the current frame The ordinate of the centroid of the target object in the image plane Indicates the number of frames in the next frame The x-coordinate of the centroid of each candidate target object in the image plane Indicates the number of frames in the next frame The ordinate of the centroid of each candidate target object in the image plane Indicates the first in the current frame The target object and the first in the next frame The Euclidean distance between the centroids of the candidate target objects in the image plane.

[0034] The process of calculating the contour intersection-union ratio is as follows: obtain the first contour in the current frame. The bounding rectangle of the target object is determined by the coordinates of its top-left corner, its width, and its height; obtain the bounding rectangle of the next frame. The bounding rectangle of the candidate target object is synthesized; the area of ​​the intersection and the area of ​​the union of the two synthesized bounding rectangles are calculated, and the area of ​​the intersection is divided by the area of ​​the union to obtain the bounding rectangle of the current frame. The target object and the first in the next frame Contour intersection ratio among candidate target objects If the area of ​​the union region is 0, then the intersection-union ratio of the contours is... Set it directly to 0.

[0035] Using the Euclidean distance of the centroid Intersection and union ratio of contours Construct the inter-frame matching cost matrix. The first... Line number Column element values It is derived from the fusion of the centroid Euclidean distance and the intersection-union ratio of the profile, and is specifically calculated using the following formula:

[0036] in, The index number of the target object in the current frame. The index number of the candidate target object in the next frame; For the current frame, the first The target object and the first in the next frame The Euclidean distance between the centroids of the candidate target objects; The length of the image plane diagonal is used as a distance normalization constant, the value of which is determined by the image resolution and is maintained. Between 0 and 1; For the current frame, the first The target object and the first in the next frame The contour intersection-union ratio among candidate target objects; This is the distance cost weighting coefficient, with a value ranging from 0 to 1. Here it is set to 0.6. The basis for this setting is: when there is a conflict between distance changes and appearance changes, spatial location continuity is trusted first, and the distance item is given a weight of 60%, while the appearance intersection-combination ratio item has a weight of 40%. Indicates the first in the current frame The target object and the first in the next frame The overall matching cost of each candidate target object; the smaller the value, the higher the degree of matching.

[0037] For all target objects in the current frame and all candidate target objects in the next frame, calculate all The values ​​are arranged to form the inter-frame matching cost matrix. If the number of target objects in the current frame is... The number of candidate target objects in the next frame is Then the dimension of the inter-frame matching cost matrix is OK List.

[0038] Perform Hungarian assignment processing on the inter-frame matching cost matrix. The input to the Hungarian assignment processing is... OK The cost matrix of the column outputs several pairing relationships. Each pairing relationship contains an index of the target object in the current frame and an index of a candidate target object in the next frame that matches it, minimizing the sum of the total costs of all pairings. The Hungarian assignment process iterates through steps such as row and column reduction, trial assignment, drawing zero lines, and adjusting the matrix to find the optimal solution. and When they are not equal, virtual rows or columns are introduced. The value of the virtual element is set to a value that is greater than any legal value, to ensure that each target object matches at most one candidate target object, and the unmatched part is automatically in an unpaired state.

[0039] Based on the optimal pairing relationship determined by the Hungarian assignment process, the centroid coordinates of the successfully paired candidate target objects in the next frame are appended to the end of the trajectory point sequence of the corresponding target object. Specifically, each target object maintains a trajectory point sequence, which stores the centroid coordinates and corresponding timestamps of the target object in each frame in timestamp order. For the first target object successfully paired in the optimal pairing relationship... The current frame target object, matching its first... The centroid coordinates of the candidate target object in the next frame The timestamp corresponding to the next frame is appended to the end of the trajectory point sequence of the target object.

[0040] When a target object cannot establish a pairing relationship with any candidate target object in the next frame within the current frame, trajectory tracking for that target object is terminated. Situations where a pairing relationship cannot be established include: the target object in the current frame being assigned to a virtual candidate target object in the Hungarian assignment process, or the cost of the target object and all candidate target objects in the matching cost matrix exceeding a predefined maximum cost threshold. After terminating tracking, the currently accumulated trajectory point sequence of the target object is output as its complete motion trajectory sequence. The motion trajectory sequence records the centroid coordinates and timestamps of the target object in each frame from the start frame to the end frame, forming a complete description of the target object's spatiotemporal displacement.

[0041] See Figure 4The figure shows the motion trajectories of three target objects obtained through temporal trajectory tracking processing in Example 3. The horizontal axis represents the pixel coordinates. The vertical axis is the pixel vertical axis. All units are pixels. The legend identifies the movement trajectories of target object 1, target object 2, and target object 3, using solid lines with dots, dashed lines with squares, and dotted lines with triangles.

[0042] The overall trajectory of target object 1 shows a trend of moving from the upper right to the lower left. Its horizontal pixel coordinate ranges from approximately 100 to 510 pixels, while its vertical coordinate decreases from approximately 200 pixels to approximately 40 pixels, indicating that the target object is moving diagonally downwards within the image plane. The trajectory lines are relatively continuous, and the dotted markers are evenly distributed, indicating that target object 1 is effectively tracked in consecutive frames.

[0043] The overall trajectory of target object 2 moves from the upper right to the lower left, but its initial ordinate is relatively high, approximately 360 pixels, and its final ordinate is approximately 150 pixels, with the horizontal coordinate ranging from 50 to 300 pixels. The trajectory is presented as a dashed line, indicating that it was drawn as a dashed line during data acquisition. The trajectory points of target object 2 are also relatively dense, indicating good continuity of its position in adjacent frames.

[0044] The trajectory of target object 3 differs significantly from the position range of target objects 1 and 2, generally exhibiting a movement trend from the lower left to the upper right. Its horizontal pixel coordinate ranges from approximately 360 to 500 pixels, while its vertical coordinate rises from approximately 0 pixels to approximately 350 pixels. The trajectory is a dotted line, and the relatively continuous trajectory line indicates that target object 3 is moving in the opposite direction within the image plane.

[0045] Example 4: In a specific implementation, the motion trajectory sequence is processed by spatial coordinate mapping, and the motion trajectory sequence is projected onto a three-dimensional spatial coordinate system constructed with the pump truck hopper opening as the center to generate a trajectory spatial distribution matrix.

[0046] Extract the pixel coordinates (horizontal and vertical) of each trajectory point in the image plane coordinate system from the motion trajectory sequence, and obtain the corresponding timestamp information. The motion trajectory sequence consists of a sequence of trajectory points of multiple target objects. Each trajectory point sequence contains trajectory points arranged chronologically. Each trajectory point records the pixel horizontal and vertical coordinates of the target object in the original image frame, as well as the video frame acquisition time. Let any trajectory point be denoted as the _th_ ... The trajectory point, the th _ ... The pixel x-coordinate value of each trajectory point is The pixel ordinate value is The timestamp information is ,in Take the index number of all trajectory points in the motion trajectory sequence.

[0047] Obtain the perspective projection transformation matrix between pre-calibrated image pixel coordinates and 3D spatial coordinates. The calibration process of the perspective projection transformation matrix is ​​as follows: At the concrete pump truck operation site, a set of 3D calibration points with known spatial positions are pre-arranged around the pump truck hopper opening. The coordinate values ​​of the 3D calibration points are measured and calibrated in a 3D spatial coordinate system constructed with the geometric center of the pump truck hopper opening as the origin, along the plane of the hopper opening and its normal direction. Simultaneously, video images containing all 3D calibration points are acquired, and the pixel abscissa and ordinate are extracted from the corresponding pixel positions of each 3D calibration point in the image. Using the correspondence between multiple pairs of 3D spatial coordinates and pixel coordinates, a 3x4 perspective projection transformation matrix is ​​calculated using a direct linear transformation algorithm. Each element in the matrix is ​​denoted as . The perspective projection transformation matrix is ​​stored in the processor's memory.

[0048] The pixel coordinates and corresponding timestamp information of each trajectory point are input into the perspective projection transformation matrix for linear transformation to obtain the three-dimensional spatial coordinates of the trajectory point in the three-dimensional coordinate system. For the first... For each trajectory point, its pixel x-coordinate value Pixel ordinate value timestamp The constant 1 is combined into a 4-dimensional column vector, which is then multiplied by the perspective projection transformation matrix to calculate the first... Three-dimensional spatial coordinates of each trajectory point The calculation formula is as follows:

[0049] in, The index number of the trajectory points in the motion trajectory sequence; For the first The pixel x-coordinate value corresponding to each trajectory point For the first The pixel ordinate values ​​corresponding to each trajectory point; These are the elements of the perspective projection transformation matrix, and the specific values ​​of each element are determined and remain unchanged through the calibration process. For the first After transformation, the trajectory points in the three-dimensional coordinate system axis coordinate values, For the first After transformation, the trajectory points in the three-dimensional coordinate system axis coordinate values, For the first The Z-axis coordinates of the trajectory points in the three-dimensional coordinate system after transformation.

[0050] Arrange all three-dimensional spatial coordinates of the same target object into a three-dimensional trajectory vector according to timestamp order, and combine the three-dimensional trajectory vectors of all target objects into a trajectory space distribution matrix. For any target object, all trajectory points it traverses are arranged according to timestamp. Arrange the points in ascending order and extract the corresponding trajectory points. The three components constitute a three-dimensional trajectory vector, with each row of the three-dimensional trajectory vector corresponding to a three-dimensional spatial position at a given time point. The three-dimensional trajectory vectors of all target objects are stacked and combined in either the row or column direction to form a trajectory space distribution matrix. Each row of the trajectory space distribution matrix corresponds to a target object, and the elements within each row are three-dimensional spatial coordinate values ​​arranged in chronological order.

[0051] In practice, based on the trajectory spatial distribution matrix, the positional overlap time series analysis of the motion trajectory of each target object and the preset hopper opening space region is performed to generate the trajectory-region overlap time series.

[0052] Extract the three-dimensional spatial coordinates of each trajectory point of each target object from the trajectory spatial distribution matrix. Let the first... The first target object The three-dimensional spatial coordinates of the trajectory points are: .

[0053] Each 3D spatial coordinate value is compared point-by-point with the boundary coordinate range of the hopper opening spatial region to determine whether the trajectory point falls within the hopper opening spatial region. The hopper opening spatial region is predefined as a cuboid spatial region with the geometric center of the pump truck hopper opening as the origin of the coordinate system. This cuboid spatial region is... The boundary range in the axial direction is ,exist The boundary range in the axial direction is ,exist The boundary range in the axial direction is The coordinate values ​​of each boundary are set according to the actual physical dimensions of the pump truck hopper and the operating area, and are stored as fixed parameters. For trajectory points... At the same time, judge Is it in Within the interval Is it in Within the interval, and Is it in Within the interval. When all three conditions are met, the trajectory point is determined to fall within the space area of ​​the hopper opening; when any one condition is not met, the trajectory point is determined not to fall within the space area of ​​the hopper opening.

[0054] For each target object, the time values ​​corresponding to the trajectory points falling into the space area inside the hopper opening are recorded sequentially according to the time order of their trajectory points, forming the initial time mark sequence corresponding to that target object. For the first... Given a target object, iterate through all its trajectory points sorted by time, and find the time values ​​corresponding to the trajectory points that meet the landing conditions. Extract and arrange the data according to the original time sequence to generate a list of time values, i.e., the initial time marker sequence.

[0055] The initial time stamp sequence performs interval merging on consecutive time values, combining time points where the difference between adjacent time points is less than the frame interval threshold into the same overlapping time interval. The frame interval threshold is set based on the video capture frame rate and is taken as 1.5 times the video frame interval time. The video frame rate is determined by reading the video stream parameters during system initialization. The video frame rate is denoted as _____. The video frame interval is Frame interval threshold Calculated as The frame interval threshold is set to 1.5 times the video frame interval because: when the target object continuously stays within the hopper opening space, the time interval between trajectory points of adjacent frames is strictly equal to the video frame interval; when there is a brief occlusion or detection jitter, there may be no more than one missed detection, causing the time interval between adjacent records to become twice the video frame interval. Setting the merging threshold to 1.5 times the video frame interval can suppress small breakpoints while avoiding the erroneous merging of independent entry events that are far apart. In the initial time marker sequence, starting from the first time value, the difference between the current time value and the next time value in the same sequence is calculated item by item. If the difference is less than... If the difference is greater than or equal to the current overlapping time interval, then the next time value will be included in the current overlapping time interval; If the current overlapping time interval ends, the current time value is taken as the end time of the interval, and the next time value is taken as the start time of the next overlapping time interval. Each overlapping time interval has a start time and an end time.

[0056] Arrange the start and end times of all overlapping time intervals in chronological order to generate a trajectory-region overlapping time series. For all the obtained overlapping time intervals, sort them in ascending order of their start times, and if the start times are the same, sort them by their end times. Output a time series containing several pairs of start and end times, i.e., the trajectory-region overlapping time series.

[0057] Example 5: In specific implementation, the trajectory-region overlap time series is processed by cumulative duration calculation to generate statistical values ​​of the residence time of each target object within the hopper opening space area. The start and end times of each overlap time interval are extracted from the trajectory-region overlap time series. The trajectory-region overlap time series consists of several overlap time interval records, each containing a start time value and an end time value, and is associated with the target object that generated this overlap time interval. For the first... Each target object corresponds to a trajectory-region overlap time series. The overlapping time intervals, the first The starting time value of each overlapping time interval is denoted as The end time value is recorded as ,in Retrieve the unique identifier index of all tracked target objects. Take all The index number of all overlapping time intervals corresponding to each target object.

[0058] For each overlapping time interval, perform a difference calculation between the end time and the start time to obtain the single dwell time corresponding to that overlapping time interval. For the th The first target object The duration of a single stay within each overlapping time interval is obtained by subtracting the start time from the end time. The durations of single stays across all overlapping time intervals corresponding to the same target object are summed to obtain the [number of]th [intervals / time intervals]. The cumulative residence time of a target object within the hopper opening space area. The calculation formula is as follows:

[0059] in, A unique identifier index for the target object; For the first The total number of overlapping time intervals contained in the trajectory-region overlapping time series of a target object; For the first The overlapping time interval index corresponding to each target object ranges from 1 to... Integers; For the first The first target object The starting time value of each overlapping time interval; For the first The first target object The end time value of each overlapping time interval; For the first The cumulative dwell time of a target object within the hopper opening space area.

[0060] The cumulative dwell time of all target objects in the trajectory-region overlap time series is summarized to generate dwell time statistics. The identifier index of all target objects and their corresponding cumulative dwell time are then used to generate these statistics. The organization is a set of key-value pairs, where the key is the target object identifier index and the value is the cumulative dwell time. This set of key-value pairs is output as a dwell time statistic.

[0061] When the cumulative dwell time exceeds the preset alarm threshold, a water-adding behavior warning signal is triggered and pushed to the on-site management terminal. Each cumulative dwell time in the statistics is compared with the alarm threshold to determine if the cumulative dwell time exceeds the threshold. The alarm threshold is preset to 30 seconds, based on the construction site safety operating procedures, which stipulate that personnel not required for construction purposes or targets carrying water-adding equipment staying in the hopper opening space area for more than 30 seconds is considered a violation of water-adding regulations. When the cumulative dwell time of a target object exceeds 30 seconds, the target object's identity and trajectory spatial coordinates within the hopper opening space area are determined at the current time. The target object's identity is determined by a unique number assigned to the target object in the target state vector set in the initial frame and is used throughout the tracking process. The trajectory spatial coordinates are extracted from the target object's 3D trajectory vector in the trajectory spatial distribution matrix by timestamp, capturing the 3D spatial coordinate sequence of all trajectory points from the moment the target object entered the hopper opening space area to the current moment.

[0062] Based on the trajectory spatial coordinates, video segments are extracted from the continuous video stream data between the moment the target object enters the hopper opening spatial region and the current moment, serving as visual evidence of the violation. The moment of entry into the hopper opening spatial region is determined by the start time of the first overlapping time interval of the target object in the trajectory-region overlap time series. During the video segment extraction process, the corresponding video frame segments are located in the continuous video stream data storage unit according to the timestamp interval, and H.264 or H.265 format encoding and compression are performed to generate independently playable video files.

[0063] The video evidence of the violation, along with the target's identification and the current timestamp, is packaged into an early warning signal and pushed to the on-site management terminal via a wireless communication link. The packaging process uses a JSON data structure, embedding the video evidence of the violation as a Base64-encoded string, recording the target's identification as a string field, and recording the current timestamp as a Unix timestamp in milliseconds, thus forming the early warning signal data packet. The wireless communication link uses a 4G wireless transmission module deployed at the pump truck operation site to send the early warning signal data packet to the on-site management terminal with a pre-set IP address via the mobile communication network. Upon receiving the early warning signal, the on-site management terminal extracts the target's identification and the video evidence of the violation, displays the alarm information on the management interface, and plays the video.

[0064] In practical implementation, when performing binarization thresholding on edge gradient components, the adaptive Otsu method is used to calculate the binarization segmentation threshold. Based on this threshold, pixels in the edge amplitude response map are divided into edge points and non-edge points. The edge amplitude response map in the edge gradient components contains the edge intensity value corresponding to each pixel location. The adaptive Otsu method uses the edge amplitude response map as input and iterates through all possible gray levels within it. For each candidate gray level, pixels with edge intensity values ​​less than or equal to that gray level are classified as background, and pixels with edge intensity values ​​greater than that gray level are classified as foreground. The mean edge intensity of background pixels, the mean edge intensity of foreground pixels, the proportion of background pixels to the total number of pixels, and the proportion of foreground pixels to the total number of pixels are calculated, and then the inter-class variance between the background and foreground classes is calculated. Among all candidate gray levels, the gray level that maximizes the inter-class variance is selected as the binarization segmentation threshold calculated by the adaptive Otsu method. The edge intensity value corresponding to each pixel position in the edge amplitude response map is compared with the binarization segmentation threshold. When the edge intensity value is greater than the binarization segmentation threshold, the corresponding pixel position is marked as an edge point; when the edge intensity value is less than or equal to the binarization segmentation threshold, the corresponding pixel position is marked as a non-edge point.

[0065] See Figure 5 In the figure, the horizontal axis represents the sequence number of overlapping events in the trajectory-region overlap time series, increasing from 0 to over 500; the left vertical axis represents the single dwell time in seconds, ranging from 0 to 25 seconds, and is displayed as a blue dot matrix; the right vertical axis represents the cumulative dwell time in seconds, ranging from 0 to 4000 seconds, and is represented by an orange solid curve; the red dashed line represents the preset duration alarm threshold of 30 seconds.

[0066] The distribution of the blue dots representing single dwell times on the vertical axis is relatively dispersed. The vast majority of single dwell times are below 5 seconds, while a few events have single dwell times reaching or approaching 25 seconds, indicating that the target object repeatedly and briefly remained in the hopper opening area. A few events showed longer single dwell times. The orange curve representing cumulative dwell time shows a steady upward trend with a relatively smooth overall trend, indicating that as the sequence number of overlapping events increases, the cumulative dwell time of the target object in the hopper opening area continues to accumulate and is generally increasing.

[0067] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for quality management of water addition during concrete construction based on deep learning, characterized in that, The method includes: Continuous video stream data of the concrete pump truck operation site is acquired, key frame extraction is performed, the image sequence to be analyzed is obtained, and spatial domain feature decomposition is performed to extract the color space component and edge gradient component of each pixel in the image. Based on the color space components and edge gradient components, the target objects in the image sequence to be analyzed are subjected to contour morphology recognition processing to generate a set of on-site object distributions. Based on the set of on-site object distribution, temporal trajectory tracking processing is performed on the object position offset between adjacent frames in the image sequence to be analyzed to obtain the motion trajectory sequence of each target object. The motion trajectory sequence is processed by spatial coordinate mapping, and the motion trajectory sequence is projected onto a three-dimensional spatial coordinate system constructed with the pump truck hopper opening as the center to generate a trajectory spatial distribution matrix; Based on the trajectory spatial distribution matrix, the positional overlap time sequence analysis of the motion trajectory of each target object and the preset hopper opening spatial region is performed to generate the trajectory-region overlap time sequence; The trajectory-region overlap time series is processed by cumulative duration calculation to generate statistical values ​​of the residence time of each target object in the hopper opening space area; When the statistical value of the dwell time exceeds the preset duration alarm threshold, a water addition behavior warning signal is triggered and the warning signal is pushed to the on-site management terminal.

2. The method for quality management of water addition in concrete construction based on deep learning according to claim 1, characterized in that, Extract the color space components and edge gradient components of each pixel in the image, including: Each frame of the image sequence to be analyzed is input into a pre-trained deep convolutional neural network, and a multi-channel feature map output from a convolutional layer in the deep convolutional neural network is extracted. The channels in the multi-channel feature map that respond to color distribution information are used as the color space components, and the channels that respond to edge gradient information are used as the edge gradient components.

3. The method for quality management of concrete construction water addition based on deep learning according to claim 1, characterized in that, The step of performing contour morphology recognition processing on target objects in the image sequence to be analyzed based on the color space components and edge gradient components to generate a set of on-site object distributions includes: The color space components and edge gradient components are input into the region proposal network to generate a set of candidate bounding boxes. The candidate target box set is processed using a classification and regression network to identify personnel targets and water supply equipment targets, and the corresponding bounding box coordinates and category confidence scores are output. Based on the bounding box coordinates, a first descriptor representing the posture features of the person and a second descriptor representing the object-carrying features of the water supply equipment are extracted; The identified personnel targets and water supply equipment targets are paired and associated according to their spatial proximity to generate a set of on-site object distributions that includes the relationships between objects carried by personnel.

4. The method for quality management of water addition in concrete construction based on deep learning according to claim 1, characterized in that, Based on the set of on-site object distributions, temporal trajectory tracking processing is performed on the object position offsets between adjacent frames in the image sequence to be analyzed to obtain the motion trajectory sequence of each target object, including: Starting from the first frame of the image sequence to be analyzed, extract the centroid coordinates and contour dimensions of each target object in the on-site object distribution set, and construct the target state vector set of the first frame; For each target object in the current frame, calculate the centroid Euclidean distance and contour intersection-union ratio between it and all candidate target objects in the next frame to form an inter-frame matching cost matrix; Perform Hungarian assignment processing on the inter-frame matching cost matrix to determine the optimal pairing relationship between the target object in the current frame and the candidate target object in the next frame; Based on the optimal pairing relationship, the centroid coordinates of the successfully paired candidate target objects in the next frame are appended to the end of the trajectory point sequence of the corresponding target object; When a target object cannot establish a pairing relationship with any candidate target object in the next frame in the current frame, the trajectory tracking of the target object is terminated, and the accumulated trajectory point sequence is output as the motion trajectory sequence of the target object.

5. The method for quality management of water addition in concrete construction based on deep learning according to claim 1, characterized in that, The step of performing spatial coordinate mapping processing on the motion trajectory sequence, projecting the motion trajectory sequence onto a three-dimensional spatial coordinate system constructed with the pump truck hopper opening as the center, and generating a trajectory spatial distribution matrix includes: Extract the pixel coordinates of each trajectory point in the motion trajectory sequence in the image plane coordinate system, and obtain the timestamp information corresponding to the trajectory point; Obtain the perspective projection transformation matrix between pre-calibrated image pixel coordinates and three-dimensional spatial coordinates; The pixel horizontal and vertical coordinates of each trajectory point and the corresponding timestamp information are input into the perspective projection transformation matrix for linear transformation processing to obtain the three-dimensional spatial coordinate values ​​of the trajectory point in the three-dimensional spatial coordinate system. Arrange all three-dimensional spatial coordinate values ​​of the same target object into a three-dimensional trajectory vector according to the timestamp order, and combine the three-dimensional trajectory vectors of all target objects into the trajectory space distribution matrix.

6. The method for quality management of water addition in concrete construction based on deep learning according to claim 1, characterized in that, Based on the trajectory spatial distribution matrix, the positional overlap time series analysis of the motion trajectory of each target object with the preset hopper opening spatial region is performed to generate a trajectory-region overlap time series, including: Extract the three-dimensional spatial coordinate values ​​corresponding to each trajectory point of each target object from the trajectory spatial distribution matrix; Each three-dimensional spatial coordinate value is compared point by point with the boundary coordinate range of the hopper opening spatial region to determine whether the trajectory point falls inside the hopper opening spatial region. For each target object, the time values ​​corresponding to the trajectory points that fall into the space area of ​​the hopper opening are recorded sequentially according to the time sequence of their trajectory points, forming the initial time mark sequence corresponding to the target object; The consecutive time values ​​in the initial time marker sequence are merged into a single overlapping time interval by merging the time points where the difference between adjacent time points is less than the frame interval threshold. Arrange the start and end times of all overlapping time intervals in chronological order to generate the trajectory-region overlapping time series.

7. The method for quality management of water addition in concrete construction based on deep learning according to claim 1, characterized in that, The process of performing a duration accumulation calculation on the trajectory-region overlap time series generates statistical values ​​of the residence time of each target object within the hopper opening space region, including: Extract the start and end time values ​​of each overlapping time interval from the trajectory-region overlapping time series; For each overlapping time interval, perform the difference calculation between the end time and the start time to obtain the single stay duration corresponding to that overlapping time interval; The cumulative residence time of the target object in the hopper opening space area is obtained by summing up the single residence time of all overlapping time intervals corresponding to the same target object. The cumulative dwell time of all target objects in the trajectory-region overlap time series is summarized to generate the dwell time statistics.

8. The method for quality management of water addition in concrete construction based on deep learning according to claim 1, characterized in that, When the statistical value of the dwell time exceeds the preset duration alarm threshold, a water addition behavior warning signal is triggered, and the warning signal is pushed to the on-site management terminal, including: The statistical value of the dwell time is compared with the duration alarm threshold to determine whether the statistical value of the dwell time is greater than the duration alarm threshold. When the statistical value of the dwell time is greater than the duration alarm threshold, the identity of the target object corresponding to the current time point and the trajectory spatial coordinates of the target object in the hopper opening space area are determined. Based on the trajectory space coordinates, a video segment is extracted from the continuous video stream data between the moment the target object enters the hopper opening space area and the current moment, as video evidence of the violation; The image evidence of the violation, along with the target object's identity and current timestamp, is packaged into a warning signal and pushed to the on-site management terminal via a wireless communication link.

9. A machine vision-based concrete construction water addition quality management system, characterized in that, The system includes a processor and a memory, the memory being electrically connected to the processor. The memory is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the memory to implement the deep learning-based concrete construction water addition quality management method as described in any one of claims 1-8.

Images