Non-visual top layer steel bar positioning device and method thereof

Abstract

The invention discloses a non-visual top reinforcing steel bar positioning device and method, and the device comprises a positioning sensor which employs a non-contact detection technology, is used for detecting the distance H between the positioning sensor and a top reinforcing steel bar in real time, and can adjust the detection angle; the detection angle is used for adapting to the spatial position of the steel bar grid; the edge calculation controller is internally provided with a UNet-based deep learning model, and the model is used for processing the image data and outputting a three-classification segmentation mask; the visual sensor is a 2D camera, is connected with the edge calculation controller and is used for collecting a steel bar grid image; wherein the positioning sensor is provided with a detection distance adjusting device, and a distance measuring threshold value is dynamically set based on the distance and the diameter of the thickest steel bar in the environment; the device cooperatively judges the intersection state of the reinforcing steel bars through primary positioning of the positioning sensor and secondary positioning of the visual sensor, and the intersection state comprises unbundled, bundled or no intersection. According to the invention, the efficiency and precision of top steel bar positioning and intersection point detection can be improved.

Description

Technical Field

[0001] This invention relates to the field of building construction technology, and in particular to a non-visual top-level rebar positioning device and method. Background Technology

[0002] In the field of building construction, multi-layered steel mesh serves as the core load-bearing foundation of concrete structures, and the precise positioning of the top layer of steel bars is a crucial prerequisite for achieving automated steel bar tying. With the acceleration of industrialization in construction, automated tying robots are gradually replacing manual operations. However, existing top-layer steel bar positioning technologies still face technical bottlenecks, particularly in balancing accuracy, cost, and environmental adaptability. Specific issues are as follows: In existing technologies, solutions for positioning top-level reinforcing bars mainly fall into two categories: One type is the high-precision positioning scheme represented by the existing technology CN 114263352 A. This scheme achieves positioning through a multi-device collaborative architecture of "depth camera + first ranging sensor + micro switch + calibration component": First, the first ranging sensor collects the distance data of the transverse rebars, triggering the electric roller to slow down; then, the micro switch (contact detection) triggers the device to stop, allowing the depth camera to align with the rebar intersection; finally, based on the 3D image collected by the depth camera, the top layer image is separated, and the intersection position is calculated by combining the world coordinates of the calibration component. Although this scheme can achieve high positioning accuracy, it has significant drawbacks: ① High hardware cost, the purchase and maintenance cost of the depth camera is much higher than that of ordinary vision equipment, making it difficult to apply on a large scale to low-cost construction scenarios; ② Complex mechanical structure, relying on the limiting groove of the electric roller and the specific support structure of the foundation, and the micro switch is a contact detection—when the rebar is deformed during construction (such as bending or displacement), it is easy to cause triggering errors, affecting positioning accuracy; ③ Difficult installation and maintenance, requiring adjustment of the roller spacing for different rebar spacings, resulting in poor adaptability.

[0003] Another type is the low-cost positioning solution represented by the existing technology CN 111780665 A. This solution uses a combination of "2D camera + line laser". By emitting a line laser to the rebar mesh, the camera captures the height difference between the rebar and the background platform (laser reflection breakpoint) to calculate the position of the rebar. Its advantage lies in low hardware cost, but its technical limitations are more prominent: ① It depends on fixed background conditions, requiring the background platform to be flat and have a stable height difference with the rebar. However, in actual construction site environments, the background platform often has problems such as debris accumulation and uneven surfaces, leading to deviations in the height difference calculation; ② It has poor adaptability to multi-layer rebar scenarios, failing to consider the interference of lower-layer rebar on the laser reflection of the top layer—lower-layer rebar will form additional reflection breakpoints, causing the 2D camera to be unable to distinguish between the top and lower layers of rebar, resulting in a significant decrease in positioning accuracy; ③ It has limited functionality, only able to locate the position of the rebar, unable to determine the binding status (bound / unbound) of the rebar intersections, requiring additional equipment for supplementary detection, reducing automation efficiency.

[0004] In summary, existing technologies suffer from high costs and poor adaptability due to reliance on depth cameras and contact detection, or insufficient accuracy and limited functionality due to reliance on background platforms and a lack of multi-layer rebar differentiation capabilities. These technologies fail to meet the demand for "low-cost, high-precision, highly adaptable, and multi-functional" top-level rebar positioning in building construction. Therefore, there is an urgent need for a technical solution that can avoid reliance on depth cameras, overcome the shortcomings of contact detection, adapt to multi-layer rebar scenarios, and simultaneously achieve rebar positioning and binding status determination. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a non-visual top-level rebar positioning device and method, which can improve the efficiency and accuracy of top-level rebar positioning and intersection detection.

[0006] To address the aforementioned technical problems, the first aspect of this invention discloses a non-visual top-level rebar positioning device, comprising: A positioning sensor is used to obtain distance information from the top layer of steel reinforcement in a non-contact detection manner; A vision sensor, which is a 2D camera, is used to acquire images of the steel mesh. An edge computing controller is communicatively connected to the positioning sensor and the vision sensor; The edge computing controller is configured as follows: a. Perform initial positioning based on the distance information obtained by the positioning sensor to determine whether the device is located in the potential rebar intersection area; b. Upon successful initial localization, the visual sensor is triggered to acquire an image, and the image is input into the built-in deep learning model for image recognition; c. Perform secondary localization based on the output of the deep learning model, wherein the output is used to identify the pixel category in the image, and the category includes untied intersections, tied intersections, or background.

[0007] As an optional implementation, the edge computing controller is further configured during initial positioning to: Based on the current distance H between the positioning sensor and the top layer of steel reinforcement and the preset diameter ΔH of the thickest steel reinforcement, a detection threshold D0 is dynamically calculated, where D0 = H + ΔH. If the distance value detected by the positioning sensor is within the range of 0 to D0, the initial positioning is considered successful.

[0008] As another optional implementation, the deep learning model is a UNet model based on an encoder-decoder structure, wherein the encoder is used to downsample the input image to extract features, and the decoder restores the image resolution and outputs a pixel-level segmentation mask through upsampling and skip connections with the encoder.

[0009] As another optional implementation, the encoder of the UNet model includes a multi-layer downsampling module, each layer of which includes at least one convolution operation and one downsampling operation; the decoder includes a multi-layer upsampling module, each layer of which includes one upsampling operation, a feature concatenation operation with the corresponding layer of the encoder, and at least one convolution operation.

[0010] As another optional implementation, the UNet model includes: The encoder consists of four convolutional blocks, and the computational logic of each convolutional block is as follows: ; in, =1,2,3,4; For encoder number Feature maps output by the layer; The input feature map is H and W, where H and W are the feature map dimensions and C is the number of channels. For the first Layer indivual The weights of the convolution kernel; For the first Layer indivual The weights of the convolution kernel; , For the first Convolutional bias term; BN for data normalization; ReLU for rectified linear function; MaxPool for... Max pooling downsampling operation; channel constraints ; The decoder consists of four transposed convolutional blocks. The computational logic of each transposed convolutional block is as follows:

[0011] ; in, For the first Feature maps output by the layer decoder; For decoder number The feature map output by the layer has a dimension of ; For encoder number Feature maps output by the layer; UpConv is Transposed convolution upsampling; Concat is a channel-level concatenation operation, resulting in a number of channels after concatenation. ; For the first Layer indivual The weights of the convolution kernel; For the first Layer indivual The weights of the convolution kernel; , For the first Layer convolution bias term; Channel constraints .

[0012] As another optional implementation, the UNet model further includes an output module for mapping the feature map output by the decoder to a three-class segmentation mask, the operation function of which is: ; in, This is the output three-class segmentation mask. This is the feature map output from the first layer of the decoder. for Convolution kernel weights, The output layer bias term is Softmax, which is a normalized exponential function. The output categories include unbundled cross points, bundled cross points, and background.

[0013] As another optional implementation, the positioning sensor is a laser rangefinder or an ultrasonic sensor, and the device further includes a bracket for mounting the positioning sensor, the bracket being configured to adjust the detection angle θ of the positioning sensor.

[0014] As another optional implementation, it also includes: A moving mechanism for driving the device to move on the reinforcing mesh; The edge computing controller is also configured to generate control commands based on the results of the secondary positioning and send them to the moving mechanism to control the device to move to the next test point.

[0015] As another optional implementation, the moving mechanism includes: Grooved wheel, the wheel surface of which has grooves that match the shape of the reinforcing bar, for moving along the longitudinal direction of the reinforcing bar; Wide wheels, which cooperate with the grooved wheels, are used to provide support and keep the device stable; A lifting mechanism is used to adjust the device's height off the ground to accommodate steel mesh with different floor heights; A lateral movement mechanism is used to drive the device to move laterally on a horizontal plane.

[0016] The second aspect of this invention discloses a non-visual top-level rebar positioning method, applied to the apparatus described in the first aspect of this invention, the method comprising the following steps: The system uses positioning sensors to scan and performs initial positioning based on the acquired distance information, thus initially identifying potential areas where steel bars intersect. Upon successful initial positioning, an image of the steel mesh in the current area is acquired using a visual sensor. The image is identified by a deep learning model within the edge computing controller, which outputs a pixel-level segmentation mask to complete secondary localization and state determination. Based on the results of the secondary positioning, the state of the rebar intersection is determined and the next action of the device is decided.

[0017] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: Compared to existing technologies, this invention utilizes a combination of initial positioning by a positioning sensor and secondary positioning by a visual sensor to collaboratively determine the status of rebar intersections (untied, tied, or no intersection). Initial positioning is performed by a non-contact sensor, quickly identifying areas where top-layer rebar may be present (using a D0 threshold to filter interference), thus narrowing the detection range. Secondary positioning is performed by a visual sensor and a UNet model, providing detailed image analysis of the initially positioned area to accurately determine the binding status of the intersections. This combination improves detection efficiency (avoiding indiscriminate analysis of the entire area by the visual sensor) while ensuring accuracy (the visual model compensates for the lack of "non-visual" information from the sensor). Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a structural schematic diagram of a non-visual top-level rebar positioning device disclosed in an embodiment of the present invention; Figure 2 This is a schematic diagram showing the comparison before and after distance measurement adjustment of a non-visual top-level rebar positioning device disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a non-visual top-level rebar positioning device after adjusting the detection angle, as disclosed in an embodiment of the present invention. Figure 4 This is a structural schematic diagram of a non-visual top-level rebar positioning device disclosed in an embodiment of the present invention in an application scenario; Figure 5 This is another structural schematic diagram of a non-visual top-level rebar positioning device disclosed in an embodiment of the present invention in an application scenario; Figure 6 This is another structural schematic diagram of a non-visual top-level rebar positioning device disclosed in an embodiment of the present invention; Figure 7 This is a flowchart illustrating the non-visual top-level rebar positioning method disclosed in an embodiment of the present invention; Figure 8 This is another flowchart illustrating the non-visual top-level rebar positioning method disclosed in this embodiment of the invention. Figure 9 This is a schematic diagram of the deep learning model process based on the encoder-decoder architecture disclosed in an embodiment of the present invention.

[0020] Figure 6 In the middle: 1. Grooved wheel; 2. Wide wheel; 3. Lifting mechanism; 4. Lateral movement mechanism; 5. Positioning sensor; 6. Positioning sensor mounting bracket; 7. Vision sensor; 8. Working platform.

[0021] Figure 2 , 4 In the diagram: H represents the distance between the positioning sensor and the top layer of reinforcing steel; ΔH represents the diameter of the thickest reinforcing steel in different environments; D0 represents the maximum threshold for positioning and identification by the positioning sensor; H r It consists of multiple layers of reinforcing bars, and the distance between each layer of reinforcing bars. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0023] Example 1 See Figures 1-6 This invention discloses a non-visual top-level rebar positioning device, comprising: Positioning sensor 5 is used to obtain distance information to the top layer of steel bars in a non-contact detection manner; Vision sensor 7 is a 2D camera used to acquire images of the steel mesh; An edge computing controller is communicatively connected to the positioning sensor and the vision sensor; The edge computing controller is configured as follows: a. Perform initial positioning based on the distance information obtained by the positioning sensor to determine whether the device is located in the potential rebar intersection area; b. Upon successful initial localization, the visual sensor is triggered to acquire an image, and the image is input into the built-in deep learning model for image recognition; c. Perform secondary localization based on the output of the deep learning model, wherein the output is used to identify the pixel category in the image, and the category includes untied intersections, tied intersections, or background.

[0024] In this embodiment of the invention, distance detection using a non-visual positioning sensor can specifically identify the distance range of the top-layer rebar, eliminating interference from lower-layer rebars in a multi-layer rebar mesh at the source, achieving "layered focusing" for initial positioning. Combined with deep learning recognition from a 2D visual sensor, the status of intersection points (untied / tied) is precisely verified a second time, ensuring positioning accuracy. Initial positioning quickly filters potential areas, while secondary positioning accurately identifies the work object; the two work together to reduce invalid detection and manual confirmation steps. Furthermore, the edge computing controller can directly communicate with construction equipment such as the tying robot, achieving an automated closed loop of "positioning → work." Replacing high-cost 3D / depth cameras with ordinary 2D cameras significantly reduces hardware procurement costs. Simultaneously, the edge computing controller integrates a deep learning model locally, eliminating reliance on high-bandwidth data transmission from the cloud, further saving computing power and maintenance costs. Non-visual positioning (such as laser ranging, ultrasonic detection, etc.) is not affected by ambient light or the surface condition of steel bars (rust, oil stains, uneven color), and can stably complete the initial positioning; 2D vision combined with deep learning models is robust to complex environments such as construction site "dust cover, uneven lighting" and can accurately identify the type of intersection.

[0025] In this embodiment, the 2D camera is a regular 2D camera, which is an imaging device used to capture two-dimensional planar images of objects.

[0026] In an optional embodiment, the edge computing controller is further configured to: Based on the current distance H between the positioning sensor and the top layer of steel reinforcement and the preset diameter ΔH of the thickest steel reinforcement, a detection threshold D0 is dynamically calculated, where D0 = H + ΔH. If the distance value detected by the positioning sensor is within the range of 0 to D0, the initial positioning is considered successful.

[0027] This embodiment dynamically generates a detection threshold D0 based on the current distance H between the positioning sensor and the top layer of reinforcing steel, and the preset diameter ΔH of the thickest reinforcing steel. If the sensor detection distance is within the range of 0 to D0, the initial positioning is considered successful. It adaptively covers the "distance fluctuation range" of the top layer of reinforcing steel, accurately eliminating interference from lower layer reinforcing steel, and improving the versatility and accuracy of the initial positioning.

[0028] For example: Assuming the top layer of rebar is 12cm from the positioning sensor (H = 12cm), and the thickest rebar used on site has a diameter ΔH = 2.5cm, then the dynamic calculation yields D0 = 12 + 2.5 = 14.5cm. In this case, if the lower layer of rebar is 25cm from the sensor, much larger than D0, the positioning sensor will only respond to the top layer of rebar within the range of 0 to 14.5cm, completely preventing the lower layer of rebar from being misidentified as the "top layer target" and ensuring the accuracy of the initial layered positioning.

[0029] In another optional embodiment, the deep learning model is a UNet model based on an encoder-decoder structure, wherein the encoder is used to downsample the input image to extract features, and the decoder restores the image resolution and outputs a pixel-level segmentation mask through upsampling and skip connections with the encoder.

[0030] In another optional embodiment, the encoder of the UNet model includes multiple downsampling modules, each of which includes at least one convolution operation and one downsampling operation; the decoder includes multiple upsampling modules, each of which includes one upsampling operation, a feature concatenation operation with the corresponding layer of the encoder, and at least one convolution operation.

[0031] In yet another optional embodiment, the UNet model includes: The encoder consists of four convolutional blocks, and the computational logic of each convolutional block is as follows: ; in, =1,2,3,4; For encoder number Feature maps output by the layer; The input feature map is H and W, where H and W are the feature map dimensions and C is the number of channels. For the first Layer indivual The weights of the convolution kernel are used for basic feature extraction; For the first Layer indivual The weights of the convolution kernel are used to optimize feature representation; , For the first Convolutional bias term is used to fine-tune the feature distribution; BN is a data normalization operation; ReLU is a linear rectified function that introduces non-linear feature selection; MaxPool is... Max pooling downsampling reduces the feature map size to half its previous size; channel constraints .

[0032] The encoder passes through 4 layers The calculation yields deep features. For the encoder's first Layer, define the operation function as .

[0033] The decoder consists of four transposed convolutional blocks. The computational logic of each transposed convolutional block is as follows:

[0034] ; in, For the first Feature maps output by the layer decoder; For decoder number The feature map output by the layer has a dimension of ; For encoder number Feature maps output by the layer (skip connection features); UpConv is Transposed convolution upsampling with a stride of 2 doubles the feature map size; Concat performs a channel-dimensional concatenation operation, resulting in a final channel count of [number missing]. ; For the first Layer indivual The weights of the convolution kernel are used to fuse and concatenate features; For the first Layer indivual The weights of the convolution kernels are used to optimize the fusion result; , For the first Layer convolution bias term; Channel constraints .

[0035] The decoder uses 4 layers The operation involves feature fusion to recover the segmentation mask. For the l-th layer of the decoder, the operation function is defined as follows: .

[0036] In another optional embodiment, the UNet model further includes an output module, which maps the feature map output by the decoder to a three-class segmentation mask. The operation logic is as follows: ; in, This is the output three-class segmentation mask. This is the feature map output from the first layer of the decoder; for Convolution kernel weights map features to three classification channels; is the output layer bias term; Softmax is the normalized exponential function; the output categories include unbundled crosspoints, bundled crosspoints, and background.

[0037] Optionally, the encoder's first... The two convolutional kernels of the layer satisfy: ,in, for Initialization function, ensure Parameter distribution and Adaptation improves the stability of feature extraction. In this encoder, each layer (the...) The second convolutional kernel of the layer The initial value of ) is not set randomly, but is based on the first convolutional kernel of that layer ( The initial value of ) is obtained through It is calculated using specific rules. In this embodiment, the encoder extracts deep features through 4 layers of convolutional blocks (including downsampling, BN, and ReLU); the decoder restores resolution and fuses details through 4 layers of transposed convolutional blocks (including upsampling, skip connections, and feature concatenation); and the final output module achieves three-class segmentation of "unbundled intersections, bound intersections, and background" through Softmax.

[0038] Deep feature extraction of the encoder enhances environmental robustness: By expanding the receptive field through multi-layer downsampling, core semantic features can be captured in complex scenes (such as steel corrosion and image blur).

[0039] Decoder skip connections + upsampling for accurate detail recovery: It fuses shallow "fine-grained details" (such as cable tie edges and steel bar textures) and deep "semantic features" from the encoder to ensure pixel-level segmentation accuracy.

[0040] The output module is categorized into three types and directly linked to construction tasks: Softmax outputs clear pixel categories, providing direct decision-making basis for equipment such as bundling robots on "whether to operate", reducing manual intervention.

[0041] Kaiming initialization optimization for convolutional kernels: the second convolutional kernel in each layer of the encoder. KaimingInit, based on the first convolutional kernel in the same layer The initial values ​​are calculated to ensure parameter distribution fit and improve feature extraction stability. This avoids gradient vanishing / exploding in the early stages of training, accelerates model convergence, and improves recognition accuracy in small samples or complex scenarios.

[0042] The following example illustrates the step-by-step computation process of the UNet model using a construction site image containing "intersections of tied rebars" (taking an input image size of 128×128×3 as an example, with 3 channels representing RGB colors): Step 1: Encoder "Downsampling to extract deep features" The encoder contains four convolutional blocks. Each layer gradually reduces the feature map size and expands the receptive field through "convolution + BN + ReLU + MaxPool" while extracting semantic features.

[0043] The first layer of convolutional blocks ( =1): Input: Original image (Height × Width × Number of Channels).

[0044] Operations: First convolution: using a 3×3 convolution kernel (Input channel 3, output channel 6), with bias applied Then, BN (batch normalization) and ReLU (linear rectification, introducing nonlinearity) are used to obtain intermediate features.

[0045] The second convolution: using a 3×3 convolution kernel Add bias , followed by BN and ReLU.

[0046] MaxPool: 2×2 max pooling (stride 2), reducing the feature map size to 64×64 while maintaining 6 channels (because...). ).

[0047] Output: At this point, the feature map has initially extracted the basic features of "rebar edge and tie outline".

[0048] 2nd layer convolution block ( =2): enter: .

[0049] Operation: Repeat "convolution + BN + ReLU + MaxPool", increasing the number of channels in the convolution kernel ( , MaxPool's size is reduced to 32×32, and the number of channels... =12.

[0050] Output: (The features are more abstract, and the semantics of "rebar intersection structure" are being captured.)

[0051] The third convolutional block ( =3): enter: .

[0052] Operations: The number of channels in the convolution kernel continues to increase ( , MaxPool's size is reduced to 16×16, and the number of channels... =24.

[0053] Output: (Features focus on the core semantics of "whether there is a cable tie at the intersection").

[0054] 4th convolutional block ( =4): enter: .

[0055] Operations: The number of channels in the convolution kernel has been upgraded again ( , MaxPool reduces the size to 8×8, and the number of channels... =48.

[0056] Output: Deep features (At this point, the features have been highly abstracted and can stably represent the global semantics of "intersections that have been bundled together").

[0057] Step 2: Decoder "Upsampling + Skip Connections, fusing details and semantics" The decoder contains four transposed convolutional blocks that gradually restore the feature map size through "upsampling + skip connections + convolution" while fusing the encoder's "shallow fine-grained details" (such as cable tie edges) with "deep semantic features" (such as intersection categories).

[0058] The first layer transposed convolutional block (corresponding to the skip connection in the fourth layer of the encoder): Input: Deep features .

[0059] Operations: UpConv: 2×2 transposed convolution (stride 2) enlarges the feature map size to 16×16 and halves the number of channels to 24 (because... , here =48), obtained .

[0060] Skip connection: Introducing the output of the encoder's third layer (Includes fine-grained details of the "cable tie edges"), and UpConv ( Perform Concat (channel concatenation) to obtain (24+24 channels).

[0061] Convolution fusion: using two 3×3 convolution kernels Add bias This is followed by BN and ReLU to optimize the fused features.

[0062] Output: (It combines "deep semantics" and "cable tie details").

[0063] Subsequent transposed convolutional blocks (repeated logic, gradually restoring size): Through a process similar to "upsampling → skip connections → convolution", the outputs of the encoder's second and first layers are sequentially introduced, and finally the output of the decoder's fourth layer is obtained. (The size is restored to 128×128 of the input image, with 6 channels).

[0064] Step 3: Output module "Three-class segmentation, output pixel-level results" The output module maps the feature map output by the decoder to three classification probabilities: "unbundled crossover, bundled crossover, and background" through 1×1 convolution.

[0065] Input: Decoder's final output .

[0066] Operations: 1×1 convolution: using a convolution kernel (Mapped to 3 category channels), with bias added This yields an unnormalized score map.

[0067] Softmax normalization: Softmax is applied to the scores of the three channels of each pixel to obtain the pixel-level classification probability.

[0068] Output: Three-class segmentation mask Y Each pixel has three channels corresponding to the probabilities of "untied intersection, tied intersection, and background".

[0069] Contextualized Results: Identifying "Bundled Intersections" Assume that in the input image, there are cable ties at the intersections of the reinforcing bars (i.e., "tied intersections"): deep features of the encoder It will highlight the overall semantic meaning of "cable tie + cross rebar"; The decoder uses skip connections to fuse the shallow "cable tie edge texture" of the encoder with deep semantics, and accurately marks the cable tie area after restoring the size; In the Softmax result of the output module, the channel corresponding to the "bundled intersection" will output a high probability (e.g., the probability of a certain pixel is [0.05, 0.90, 0.05], which corresponds to "unbundled, bundled, and background" respectively), thereby achieving pixel-level accurate classification and providing the bundling robot with a decision basis for "no need for repetitive work".

[0070] Through the above steps, the UNet model completes the end-to-end computation from the "original image" to the "pixel-level classification mask", which ensures the robustness of deep semantics and preserves the accuracy of fine-grained details.

[0071] In another optional embodiment, the positioning sensor is a laser rangefinder or an ultrasonic sensor, mounted on a positioning sensor mounting bracket, which is adapted to the sensor angle adjustment requirements.

[0072] The positioning sensors utilize either laser rangefinders or ultrasonic sensors, both of which are non-contact detection devices. Laser rangefinders offer high accuracy (typically ±1mm) and strong directionality, enabling precise focusing on top-level rebar even under complex lighting conditions (such as strong sunlight or shadows at construction sites), avoiding interference from side rebars or the background. This makes them suitable for scenarios with stringent positioning accuracy requirements. Ultrasonic sensors offer strong anti-interference capabilities (good tolerance to dust and slight obstruction) and relatively low cost. They can operate stably in harsh environments with high dust and humidity, making them suitable for large-area, high-speed scanning scenarios.

[0073] The sensor is mounted on a positioning sensor mounting bracket, which adapts to the adjustment requirements of the sensor's detection angle θ via a mechanical or electric structure: For a mechanical adjustment bracket, it features a damped rotary joint or a graduated angle adjustment seat, allowing the operator to manually rotate the sensor and adjust the detection angle θ to the target value (e.g., 0° vertical, 15° tilt, etc.) using the graduated markings. This is convenient and allows for real-time angle locking. For an electric adjustment bracket, it is equipped with a rotation module driven by a stepper motor or servo motor, enabling automatic angle adjustment via program control (e.g., receiving controller commands to adjust θ from 0° to 30°), suitable for scenarios requiring high automation and remote / batch angle adjustments.

[0074] In yet another optional embodiment, it further includes: A moving mechanism for driving the device to move on the reinforcing mesh; The edge computing controller is also configured to generate control commands based on the results of the secondary positioning and send them to the moving mechanism to control the device to move to the next test point.

[0075] In yet another alternative embodiment, the moving mechanism includes: Grooved wheel 1 has a groove on its surface that matches the shape of the reinforcing bar, for moving longitudinally along the reinforcing bar; The wide wheel 2, in conjunction with the grooved wheel 1, is used to provide support and maintain the stability of the device; Lifting mechanism 3 is used to adjust the height of the device off the ground to accommodate steel mesh with different floor heights; The lateral movement mechanism 4 is used to drive the device to move laterally on the horizontal plane.

[0076] In another optional embodiment, the grooved wheel 1 has a limiting groove on its surface, which is adapted to the diameter of the reinforcing bar to prevent it from shifting during movement.

[0077] Example 2 See Figures 2-6 In this scenario, it is necessary to locate the top layer of steel reinforcement in a two-layer steel mesh. Assume that the distance H between the top layer of steel reinforcement and the positioning sensor is 15cm, the diameter ΔH of the thickest steel reinforcement in the environment is 2cm, and the spacing between each layer of steel reinforcement is Hr.

[0078] See Figure 7 This invention discloses a non-visual top-level rebar positioning method, used in the non-visual top-level rebar positioning device as described in Embodiment 1, comprising the following steps: S1. Scanning is performed using positioning sensors, and initial positioning is completed based on the acquired distance information, thus initially screening out potential areas where steel bars intersect.

[0079] Preliminary work before step S1: First, place the device on two layers of steel mesh, with the grooved wheel pressing down on a single steel bar. Then, initialize the edge computing controller, load the preset configuration file and operating parameters, and check the controller's hardware connection status to ensure all interfaces are working properly. Next, initialize the positioning sensor and check its operating status, including but not limited to self-test programs and communication link tests, to ensure the positioning sensor establishes communication with the edge computing controller. Set the sensor's operating mode to continuous scanning to achieve real-time positioning of the top layer of steel bars. Then, adjust the detection threshold: using a mechanical or electric adjustment mechanism in the positioning sensor, based on the distance H between the top layer of steel bars in the environment and the positioning sensor being 15cm, considering the steel bar diameter ΔH being 2cm, and according to D0 = H + ΔH, and based on the detection angle θ being 0°, i.e., the positioning sensor is perpendicular to the steel bar, finally set the positioning sensor's distance threshold to 0~17cm, i.e., D0 is 17cm.

[0080] Optionally, the step of scanning with a positioning sensor and completing the initial positioning based on the acquired distance information includes: S11. Obtain the current distance between the sensor and the target object.

[0081] S12. Compare the distance value with the dynamic detection threshold D0.

[0082] The D0 is calculated using the formula D0 = H + ΔH, where H is the measured distance from the sensor to the top layer of steel reinforcement, and ΔH is the diameter of the thickest steel reinforcement in the environment.

[0083] S13. If the distance value is within the range of 0 to D0, then the initial positioning success signal is triggered.

[0084] If the distance between the sensor and the rebar is detected to be within 0-17cm, the positioning is considered successful, and the process proceeds to step S2. If the distance between the sensor and the rebar is detected to be outside the 0-17cm range, the positioning is considered unsuccessful. In this case, the scanning steps are repeated multiple times. If the scanning steps are still unsuccessful after multiple repetitions, the process proceeds to step S5, where the system will issue a warning and record the information of this failure for subsequent analysis and improvement.

[0085] S2. When the initial positioning is successful, the image of the steel mesh in the current area is acquired through the visual sensor.

[0086] After the initial positioning of the top layer of rebar, the edge computing controller receives a trigger signal from the positioning sensor and generates corresponding control commands accordingly. After the initial positioning is completed, the edge computing controller collects information through a vision sensor, which is a 2D camera.

[0087] S3. The image is identified by a deep learning model within the edge computing controller, and a pixel-level segmentation mask is output to complete secondary localization and state determination.

[0088] The image input is based on the UNet deep learning model, specifically the UNet model in Example 1. For instance, given an image of a cross-shaped rebar intersection, the UNet model performs pixel-level segmentation, and the output segmentation mask shows that the central area of ​​the intersection is identified as "untied" (highlighted), while the surrounding rebar area is identified as "background." This process is called secondary localization, which not only confirms the precise location (pixel coordinates) of the intersection but also determines its state.

[0089] S4. Based on the results of the secondary positioning, determine the state of the rebar intersection and decide on the next action of the device.

[0090] Optionally, the next step of the determining device includes: S41. If an untied intersection is identified, the control device performs a tying operation or records the location of the point; S42. If the intersection is identified as already bound or does not exist, the control mechanism moves the device to the next positioning point and repeats the above positioning process.

[0091] S5. If the positioning process fails to complete successfully, the system will stop the current operation, issue a warning, and record all diagnostic information, including sensor readings, environmental parameters, etc., and the positioning will end.

[0092] This invention employs a step-by-step strategy of "initial positioning by a positioning sensor + secondary positioning by a visual sensor." Initial positioning quickly filters potential rebar intersection areas, reducing invalid detection ranges. Secondary positioning, based on a deep learning model, achieves pixel-level segmentation, accurately identifying intersection states and significantly improving positioning accuracy while avoiding the limitations of a single sensor. Combining distance information from a non-contact positioning sensor with semantic features of 2D visual images, multi-source data collaboration compensates for the lack of depth information in pure visual methods, ensuring accurate differentiation of top-level rebar in multi-layered rebar meshes and avoiding positioning deviations due to insufficient spatial information, thus solving the problem of missing depth information. The positioning sensor can dynamically adapt to different floor heights (H) via a detection distance adjustment device. The diameter (△H) and spacing of the reinforcing bars are unaffected by ambient light, bar color, texture, or size. A standard 2D vision sensor, combined with the UNet model, eliminates the need for a depth camera, maintaining stable performance even in complex construction scenarios and is suitable for various reinforcing bar mesh layouts. The combination of a non-contact positioning sensor and a standard 2D vision sensor replaces the expensive depth camera solution, significantly reducing hardware costs. Simultaneously, real-time data processing and decision-making are achieved through an edge computing controller, eliminating the need for cloud computing and reducing deployment costs. Secondary positioning can directly output the status judgment of "untied intersection / tied intersection / background," providing a clear basis for the device's next action, reducing manual intervention, improving the automation efficiency of reinforcing bar binding, and shortening the construction cycle.

[0093] Figure 8 The diagram shown is a flowchart for locating and identifying the intersection points of the top layer of the steel mesh, briefly illustrating the core process: Initial preparation: Place the device on the steel mesh (so that the grooved wheel rests on top of the steel bars), initialize the positioning sensor and edge computing controller, and set the detection threshold according to the steel mesh parameters; Location and identification loop: First, determine the "number of times the positioning sensor is triggered": if the number of triggers is excessive, record the positioning failure parameters; if the number of triggers is not excessive, then determine whether the "initial positioning was successful" - if it fails, retry positioning by "forward and backward + left and right movement device"; if it succeeds, the edge computing controller processes the trigger signal, the vision sensor collects the intersection image, and inputs it into the UNet model to identify the intersection category (such as untied or tied). End judgment: After identification, it is determined whether the "last row / column" of the steel mesh has been reached. If not, the moving device enters the next positioning cycle; if it has, the process ends.

[0094] The overall system embodies the automated workflow of "positioning → visual recognition → movement loop → completion of the entire area".

[0095] Figure 9 The diagram shown is a flowchart of a deep learning network based on an encoder-decoder structure, which briefly explains the core process: Input and Encoder: After the 2D image is input, it is processed by an encoder consisting of “convolutional layer 1 → convolutional layer 2 → convolutional layer 3 → convolutional layer 4” to gradually extract the deep semantic features of the image (such as the abstract concepts of “intersecting steel bars and the presence of cable ties”), and finally obtains the most condensed semantic expression in the “bottleneck layer”.

[0096] Skip connections: Represented by dashed lines, these connections directly pass the fine-grained features (such as cable tie edges, steel bar textures, etc.) preserved by each convolutional layer of the encoder to the corresponding layer of the decoder, thus compensating for the problem of "detail loss due to encoder downsampling".

[0097] Decoder and output: The decoder, consisting of “upsampling block 1 → upsampling block 2 → upsampling block 3 → upsampling block 4”, restores the image resolution through upsampling and fuses “details from skip connections” with “semantics from the encoder”; finally, it is processed by an “activation function” (such as Softmax) to output pixel-level classification results (such as distinguishing “unbound intersections, bound intersections, and background”).

[0098] The overall logic is to achieve precise pixel-level analysis of images by "encoding semantics, decoding detail, and merging the two through skip connections".

[0099] The content disclosed in the embodiments of this invention is only a preferred embodiment of the invention and is used only to illustrate the technical solutions of the invention, not to limit it. Although the invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this invention.

Claims

1. A non-visual top-level rebar positioning device, characterized in that, include: A positioning sensor is used to obtain distance information from the top layer of steel reinforcement in a non-contact detection manner; A vision sensor, which is a 2D camera, is used to acquire images of the steel mesh. An edge computing controller is communicatively connected to the positioning sensor and the vision sensor; The edge computing controller is configured as follows: a. Perform initial positioning based on the distance information obtained by the positioning sensor to determine whether the device is located in the potential rebar intersection area; b. Upon successful initial localization, the visual sensor is triggered to acquire an image, and the image is input into the built-in deep learning model for image recognition; c. Perform secondary localization based on the output of the deep learning model, wherein the output is used to identify the pixel category in the image, and the category includes untied intersections, tied intersections, or background.

2. The apparatus according to claim 1, characterized in that, The edge computing controller is further configured to perform the initial positioning as follows: Based on the current distance H between the positioning sensor and the top layer of steel reinforcement and the preset diameter ΔH of the thickest steel reinforcement, a detection threshold D0 is dynamically calculated, where D0 = H + ΔH. If the distance value detected by the positioning sensor is within the range of 0 to D0, the initial positioning is considered successful.

3. The apparatus according to claim 1, characterized in that, The deep learning model is a UNet model based on an encoder-decoder structure. The encoder is used to downsample the input image to extract features, and the decoder restores the image resolution and outputs a pixel-level segmentation mask through upsampling and skip connections with the encoder.

4. The apparatus according to claim 3, characterized in that, The encoder of the UNet model includes multiple downsampling modules, each of which includes at least one convolution operation and one downsampling operation; the decoder includes multiple upsampling modules, each of which includes one upsampling operation, a feature concatenation operation with the corresponding layer of the encoder, and at least one convolution operation.

5. The apparatus according to claim 3, characterized in that, The UNet model includes: The encoder consists of four convolutional blocks, and the computational logic of each convolutional block is as follows: ； in, =1,2,3,4; For encoder number Feature maps output by the layer; The input feature map is H and W, where H and W are the feature map dimensions and C is the number of channels. For the first Layer indivual The weights of the convolution kernel; For the first Layer indivual The weights of the convolution kernel; , For the first Convolutional bias term; BN for data normalization; ReLU for rectified linear function; MaxPool for... Max pooling downsampling operation; channel constraints ; The decoder consists of four transposed convolutional blocks. The computational logic of each transposed convolutional block is as follows: ； in, For the first Feature maps output by the layer decoder; For decoder number The feature map output by the layer has a dimension of ; For encoder number Feature maps output by the layer; UpConv is Transposed convolution upsampling; Concat is a channel-level concatenation operation, resulting in a number of channels after concatenation. ; For the first Layer indivual The weights of the convolution kernel; For the first Layer indivual The weights of the convolution kernel; , For the first Layer convolution bias term; Channel constraints .

6. The apparatus according to claim 5, characterized in that, The UNet model also includes an output module, which maps the feature map output by the decoder to a three-class segmentation mask. Its operation function is as follows: ； in, This is the output three-class segmentation mask. This is the feature map output from the first layer of the decoder. for Convolution kernel weights, The output layer bias term is Softmax, which is a normalized exponential function. The output categories include unbundled cross points, bundled cross points, and background.

7. The apparatus according to claim 1, characterized in that, The positioning sensor is a laser rangefinder or an ultrasonic sensor, and the device further includes a bracket for mounting the positioning sensor, the bracket being configured to adjust the detection angle θ of the positioning sensor.

8. The apparatus according to claim 1, characterized in that, Also includes: A moving mechanism for driving the device to move on the reinforcing mesh; The edge computing controller is also configured to generate control commands based on the results of the secondary positioning and send them to the moving mechanism to control the device to move to the next test point.

9. The apparatus according to claim 8, characterized in that, The mobile mechanism includes: Grooved wheel, the wheel surface of which has grooves that match the shape of the reinforcing bar, for moving along the longitudinal direction of the reinforcing bar; Wide wheels, which cooperate with the grooved wheels, are used to provide support and keep the device stable; A lifting mechanism is used to adjust the device's height off the ground to accommodate steel mesh with different floor heights; A lateral movement mechanism is used to drive the device to move laterally on a horizontal plane.

10. A non-visual top-level rebar positioning method, applied to the device as described in any one of claims 1-9, characterized in that, The method includes the following steps: The system uses positioning sensors to scan and performs initial positioning based on the acquired distance information, thus initially identifying potential areas where steel bars intersect. Upon successful initial positioning, an image of the steel mesh in the current area is acquired using a visual sensor. The image is identified by a deep learning model within the edge computing controller, which outputs a pixel-level segmentation mask to complete secondary localization and state determination. Based on the results of the secondary positioning, the state of the rebar intersection is determined and the next action of the device is decided.

Images