A multi-modal perception poultry body size and weight intelligent measurement management method

Abstract

The application discloses a kind of multi-modal perception's poultry body weight intelligent measurement management method, it is related to three-dimensional point cloud image technical field.The present application is combined with DMB-PointNeXt segmentation model by seed constraint neighborhood search and connectivity consistency check, solve the point cloud background mixed, low foot leakage and regional boundary fuzzy problem, realize keel, chest and tarsometatarsal region accurate positioning;Based on segmented region respectively extract straight keel length and arc length keel length, front chest width and last chest width and chest length, shank length, solve the strong subjectivity of artificial measurement, the problem of poor reproducibility;Through the improvement TOPSIS score, GMM probability stratification and D-S evidence fusion generation management decision, overcome the single-index evaluation limitation and the defect of fixed threshold adaptability, significantly improve the automation level and reliability of poultry phenotype determination.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional point cloud image technology, and in particular to a multimodal sensing intelligent measurement and management method for poultry body size and weight. Background Technology

[0002] Keel length, chest length, chest width, and tibia length are important phenotypic parameters reflecting an individual's skeletal development, body structure, growth status, and breeding value. They are significant in poultry breeding, growth monitoring, and body shape assessment. Keel length is closely related to chest bone development, chest length and chest width can be used to evaluate chest development levels, and tibia length reflects tarsometatarsal bone development and lower limb support. Therefore, these parameters have high application value in poultry production management and precision breeding.

[0003] Currently, the aforementioned poultry body size parameters are still mostly measured manually in production and research. In practice, workers typically need to individually grasp and fix each poultry individual, part the feathers on the chest and abdomen to expose the tarsometatarsus-toe-claw area, and then use rulers, calipers, flexible measuring tapes, or specialized measuring tools to take readings for different parts. While this method is intuitive, it relies heavily on the experience of the measuring personnel. Different personnel may not be able to completely agree on the method of fixing the poultry, the selection of measuring points, the tightness of the measuring tape, and the reading technique, easily leading to deviations in the measurement start and end points and poor repeatability of results. In particular, live poultry are prone to struggling, retracting their legs, and twisting their bodies during fixation and measurement, often requiring repeated adjustments to their posture or even repeated measurements, thus increasing human error and operational burden. Furthermore, frequent grasping, fixing, and measuring can easily cause stress in individual poultry, which is not conducive to large-scale, continuous, and standardized body size measurement.

[0004] With the development of machine vision and intelligent sensing technologies, some publicly available solutions have begun to explore the use of images, point clouds, or integrated measurement devices for poultry body size measurement. Invention patent (CN103892837B) discloses a method for measuring the body size of waterfowl, such as keel length, chest width, and shank length, using side views, top views, and reference scale conversion. This method enables image-based body size measurement, but primarily relies on two-dimensional image contours and scale conversion, without addressing issues such as background stabilization separation of poultry body surface point clouds and point cloud boundary correction under feather interference. Furthermore, related research focuses more on local indicators related to the tarsometatarsal segment, with limited utilization of the spatial geometric features of the poultry body surface trunk region, particularly the keel region. The invention patent (CN117053875B) discloses an intelligent poultry phenotyping device and method. The device integrates components such as a weighing scale, a camera mechanism, and an RFID reader, focusing more on the structure of the measuring station and the acquisition of multi-source phenotyping information. It provides less detail on specific processing methods such as effective point cloud extraction from the poultry body surface, measurement-guided point cloud noise reduction, local envelope correction under feather influence, and fine extraction of body size parameters. Regarding 3D perception, some studies have utilized 3D vision for poultry weight prediction, but these focus more on weight prediction or general body size estimation. They do not adequately consider issues such as the stable separation of poultry body surface point clouds under fixed working conditions, local geometric processing of the chest-abdomen-wing root composite region and the keel region, and the stability determination of the anterior and posterior ends of the keel in live poultry.

[0005] Therefore, although existing technologies have explored aspects such as poultry image body size measurement, local tarsometatarsal segment index measurement, comprehensive phenotypic acquisition devices, and three-dimensional visual estimation of poultry body surface, several technical problems remain unresolved for live poultry body size measurement scenarios at fixed workstations: First, poultry body surface point clouds are easily mixed with the operating table, background, and fixed structures, affecting the stable extraction of effective point clouds from the poultry body surface; second, feather coverage on the poultry body surface can cause point cloud boundary jitter, local redundancy, and unclear real body surface boundaries, making it difficult to stably determine the measurement benchmarks for the keel region, the chest-abdomen-wing root complex region, and the tarsometatarsal-to-claw complex region; third, as an important indicator related to the development of the thoracic skeleton, the anterior and posterior endpoints of the keel are not easily and stably identified in two-dimensional images or general body size estimation methods, and there is still a lack of technical solutions for stable endpoint positioning and length measurement based on the spatial geometric features of the keel surface. Summary of the Invention

[0006] The purpose of this invention is to provide a multimodal sensing intelligent measurement and management method for poultry body size and weight, in order to solve the technical problem of how to stably extract effective data of poultry body surface from a three-dimensional point cloud that is affected by feather interference and background mixing in a single-view measurement scenario of live poultry in a fixed position, and to accurately locate anatomical measurement benchmarks such as the anterior and posterior ends of the keel, so as to realize the automated and standardized measurement of core body size parameters such as keel length, chest length, chest width and tibia length.

[0007] To achieve the above-mentioned objectives, the embodiments of the present invention provide the following technical solutions:

[0008] A multimodal sensing intelligent measurement and management method for poultry body size and weight, comprising:

[0009] A multimodal phenotypic data acquisition device for poultry was set up, and basic vital signs data and RGB point cloud data of poultry were collected; the RGB point cloud data included RGB images, depth images and raw color point cloud data;

[0010] Based on RGB point cloud data, seed-constrained neighborhood retrieval and connectivity consistency verification are used to segment the data using the DMB-PointNeXt segmentation model, generating point cloud data for different regions of poultry. The point cloud data for different regions of poultry includes point clouds of the keel region, point clouds of the chest and abdomen-wing root composite region, point clouds of the tarsus-toe and claw composite region, and point clouds of other regions.

[0011] Based on point cloud data of different regions of poultry, the keel length parameter, chest body size parameter, and tibia length are calculated; the keel length parameter includes the centerline arc length and straight keel length; the chest body size parameter includes the anterior chest width, posterior chest width, and chest length.

[0012] Based on basic physical characteristics data, keel length parameters, chest body size parameters, and tibia length, the corresponding comprehensive development score is calculated, and poultry management results are generated by combining the GMM model and DS evidence theory.

[0013] In the above process, this invention, based on RGB point cloud data, employs seed-constrained neighborhood retrieval and connectivity consistency verification, combined with the DMB-PointNeXt segmentation model to perform precise point-level semantic segmentation of the keel, thorax-wing root composite region, and tarsometatarsus-toe composite region. This effectively solves the problems of point cloud mixing with the background, easy omission of low-position slender structures, and blurred boundaries between different regions, providing accurate regional positioning for body size measurement. Furthermore, the straight keel length and centerline arc length are extracted from the keel region point cloud to meet the needs of both caliper-style straight-line measurement and keel path measurement. The thorax-wing root composite region point cloud is located using anterior chest feature points, posterior edge local corner positioning, and the front baseline. The system obtains the forechestal width, distal thoracic width, and thoracic length through constraints. The tibia length is obtained from the point cloud of the tarsometatarsus-toe complex region by locating the palmotodontical transition point and proximal bony prominence. This achieves automated and standardized extraction of multiple core body size parameters, solving the problems of reliance on experience, strong subjectivity of measurement points, and poor repeatability in manual contact measurements. By improving the TOPSIS comprehensive developmental score, GMM probabilistic state stratification, and DS evidence, poultry management results are generated. This solves the problems that single body size or weight indicators cannot fully reflect developmental status and that fixed thresholds are not adapted to differences between different batches. It significantly improves the efficiency, objectivity, and repeatability of poultry phenotyping, providing reliable technical support for poultry breeding screening, health monitoring, and precision feeding. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of the method in an embodiment of the present invention;

[0016] Figure 2 This is a structural diagram of the poultry multimodal phenotyping equipment in an embodiment of the present invention;

[0017] The components include: 1. RFID reading module at the entrance; 2. Weighing platform; 3. Temperature detection unit; 4. Depth camera; 5. Industrial all-in-one machine; 6. Power bank; 7. Operating table; 8. Vertical support structure; and 9. LED fill light. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This embodiment provides a multimodal sensing intelligent measurement and management method for poultry body size and weight. Figure 1 The execution entity of the method shown can be a software and / or hardware device. The execution entity of this application can include, but is not limited to, at least one of the following: user equipment, network equipment, etc. User equipment can include, but is not limited to, computers, smartphones, personal digital assistants (PDAs), and the aforementioned electronic devices. Network equipment can include, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of computers or network servers. Cloud computing is a type of distributed computing, consisting of a super virtual computer composed of a group of loosely coupled computers. This embodiment does not impose any limitations on this.

[0020] A multimodal sensing-based intelligent measurement and management method for poultry body size and weight includes:

[0021] S1. Set up a poultry multimodal phenotypic data acquisition device and collect basic vital signs data and RGB point cloud data of poultry; the basic vital signs data includes individual identification, weight data and rectal temperature; the RGB point cloud data includes RGB images, depth images and raw color point cloud data;

[0022] It needs to be explained that, such as Figure 2 As shown, a poultry multimodal phenotypic data acquisition device includes an inlet RFID reading module 1, a weight weighing platform 2, a temperature detection unit 3, a depth camera 4, an industrial all-in-one machine 5, a mobile power supply 6, an operating table 7, a vertical support 8, and an LED supplementary light 9.

[0023] The depth camera 4 (used to collect RGB point cloud data of poultry) is installed on the top of the vertical support 8; the LED supplementary light 9 is installed on the top of the vertical support 8, and is deployed in conjunction with the depth camera 4 to form a stable and consistent single-view visual acquisition field of view and lighting conditions, so that identification, weight measurement, temperature detection and visual acquisition can be completed sequentially at the same workstation; the industrial all-in-one machine 5 is embedded in the middle of the vertical support 8 and is communicatively connected to the depth camera 4; the entrance RFID reading module 1 (used to read individual electronic tags of poultry to complete identification) is set on the operating table 7. The front identification position is connected to the industrial all-in-one machine 5; the weight weighing platform 2 (used to collect the weight data of individual poultry) is set on one side of the operating platform 7 and is connected to the industrial all-in-one machine 5; the temperature detection unit 3 (used to measure the rectal temperature of individual poultry) is set near the working area of ​​the operating platform 7 and is connected to the industrial all-in-one machine 5; the mobile power supply 6 (used to power the entire equipment) is set at the lower part of the operating platform 7; the vertical support body 8 is fixedly set on one side of the operating platform 7, and the depth camera 4, the industrial all-in-one machine 5 and the LED fill light 9 are integrated on the vertical support body 8.

[0024] The vertical support 8 is a cabinet-style integrated support unit, forming a fixed single-view visual acquisition terminal facing the work area of ​​the operating table 7 together with the depth camera 4, LED supplementary light 9, and industrial all-in-one machine 5. The entrance RFID reading module 1, weight weighing station 2, temperature detection unit 3, and depth camera 4 are arranged around the operating table 7, enabling identity recognition, weight weighing, temperature detection, and visual acquisition to be completed sequentially at the same workstation. The depth camera 4 is connected to the industrial all-in-one machine 5 via a USB communication interface to collect RGB point cloud data of poultry; the entrance RFID reading module 1 and weight weighing station 2 are connected to the industrial all-in-one machine 5 via a serial communication interface; the temperature detection unit 3 and LED supplementary light 9 are connected to the industrial all-in-one machine 5 via an RS485 bus and use the Modbus RTU protocol for data interaction and control.

[0025] The industrial all-in-one machine 5 has a built-in client program for unified scheduling of each acquisition unit. After completing the identification of individual poultry, weight acquisition and temperature measurement, it triggers the depth camera 4 to collect data from the poultry body surface at the acquisition position, and associates, stores and displays the data of each modality according to the individual number.

[0026] The processing procedure of the poultry multimodal phenotyping equipment is as follows:

[0027] Operators individually pick up the poultry to be tested and bring each individual poultry close to the RFID reader module 1 at the entrance for individual identification. The industrial all-in-one machine 5 receives and records the RFID number of the current poultry individual as the individual identification identifier for the current data collection task. After obtaining the RFID number of the current poultry individual, the client program built into the industrial all-in-one machine 5 stops RFID monitoring and releases the corresponding communication channel. When a new RFID number is detected, the collection status, part cache information, and interface display status of the previous poultry individual are reset, and a new individual data collection task and original data directory are created according to the current RFID number.

[0028] After the individual poultry is identified, the operator places the poultry on the weighing platform 2 to start the weight collection process. The client program built into the industrial all-in-one machine 5 first performs zero-point calibration on the weighing platform 2, then continuously samples the weight signal within the preset sampling time window, and determines the stable weight state based on the fluctuation range of the sampling results, and extracts the stable weight value as the current poultry's weight data.

[0029] After the weight is collected, the rectal temperature of each poultry individual is measured through the temperature detection unit 3, and the rectal temperature, RFID number, and weight data are written into the current basic vital signs data record of the poultry.

[0030] After completing the basic vital signs collection, the operator transfers the poultry to the operating table 7 and uses flexible straps to bind and fix the poultry's feet to the operating table 7, keeping the poultry in a supine restraint position. The position and orientation of the poultry on the operating table 7 are adjusted so that the poultry's chest and abdomen face upwards, and the keel area, chest and abdomen area, and tarsometatarsus-toe-claw complex area are within the effective acquisition field of view of the depth camera 4. The LED supplementary light 9 is turned on to maintain stable and consistent lighting conditions in the acquisition area, completing the preparation for single-view visual acquisition.

[0031] After individual poultry are restrained and placed in the data collection area, the depth camera 4 is triggered by the built-in client program of the industrial all-in-one machine 5 to perform single-view synchronous data acquisition, obtaining the RGB point cloud data of the current poultry. The industrial all-in-one machine 5 binds the acquired RGB point cloud data with the corresponding basic vital signs data and acquisition time, and establishes corresponding raw data records according to the individual number. After completing the raw data acquisition of the current poultry, the industrial all-in-one machine 5 saves and displays the current acquisition results, and uses the RGB point cloud data as input data for the poultry body surface point cloud processing and body size parameter extraction process.

[0032] S2. Based on RGB point cloud data, segmentation is performed using seed-constrained neighborhood retrieval and connectivity consistency verification, combined with the DMB-PointNeXt segmentation model, to generate point clouds of the keel region, the thorax-abdomen-wing root composite region, the tarsometatarsus-toe claw composite region, and other regions.

[0033] S2 includes:

[0034] S2-1. The RGB point cloud data is processed using the Mask2Former model and the DBSCAN density connectivity clustering method. Combined with seed constraint neighborhood retrieval and connectivity consistency verification, effective point clouds of poultry body surface are generated.

[0035] To address the issue that during poultry body surface sampling in a supine restraint state, poultry body surface point clouds are easily mixed with the operating table 7, fixed structures, and background noise, and that the slender, low-lying foot structure of the poultry body surface is close to the operating table 7, making it prone to accidental deletion of point cloud data during background removal and subject selection, this application employs a method combining RGB poultry body surface mask generation, initial poultry body surface point cloud extraction, poultry body surface subject point cloud selection, low-lying structure candidate point recall, and residual noise block removal. This method removes background interference while preserving the poultry body surface contour boundary and the low-lying foot point cloud required for tibia length measurement.

[0036] S2-1 includes:

[0037] S2-1-1. Use the SAM segmentation model (Segment Anything Model) to perform auxiliary labeling on historical RGB images, generate poultry body surface segmentation labeling data, and input it into the original Mask2Former model for training to generate the Mask2Former model.

[0038] Specifically, historical RGB images covering different poultry individuals, body sizes, and feather colors were selected as annotation samples. First, the SAM segmentation model was used to perform auxiliary annotation on the poultry body surface regions. After manual verification, poultry body surface segmentation annotation data was generated and input into the original Mask2Former model for training, generating the Mask2Former model. During training, preset categories included poultry body surface categories and background categories. The trained model was used to identify poultry body surface regions from RGB images at a fixed workstation.

[0039] S2-1-2. Input the RGB image into the Mask2Former model, combine it with the original color point cloud data to generate the target poultry body surface mask and the initial poultry body surface point cloud;

[0040] Specifically, the Mask2Former model comprises a backbone network, a pixel decoder (PD), and a transformer decoder (TD). The backbone network extracts multi-scale image features rich in semantic information from the RGB image. Subsequently, the pixel decoder (PD) upsamples and fuses these multi-scale image features to restore the original image resolution, resulting in pixel-level high-resolution features where each pixel location possesses semantic representation capabilities. Building upon this, the transformer decoder (TD) internally maintains a set of learnable object queries. Each object query acts as a "probe" responsible for searching for a specific target in the image, interacting with the pixel-level features through attention calculations, thereby outputting a candidate mask for each query. (i.e., the pixel regions that the query considers to belong to a certain target) and their corresponding category probability vectors Category probability vector Indicates candidate mask The predicted probabilities belong to the two categories of "poultry body surface" and "background".

[0041] Read all category probability vectors "Poultry body surface" category Corresponding category probability vector Then, among all object queries, the object query with the highest category probability vector is selected, and its corresponding candidate mask is determined as the target poultry body surface mask. The corresponding formula is:

[0042] ;

[0043] in, Both j represent the query number of the object with the largest category probability vector. Indicates the first One candidate mask, This indicates the operation of selecting the maximum value. This represents the total number of object queries involved in the prediction process in the transformer decoder.

[0044] The relationship between the RGB image and the original color point cloud was established based on the factory calibration parameters of the depth camera. Each 3D point is associated with a pixel position in the RGB image. Extract pixel position Falling into the target poultry's body mask All three-dimensional points within the cloud form the initial point cloud of the poultry's body surface. .

[0045] S2-1-3. The initial poultry body surface point cloud is processed by RANSAC reference plane fitting, region of interest (ROI) screening and DBSCAN density connectivity clustering to generate the main body point cloud of poultry body surface.

[0046] Initial poultry body surface point cloud It may still contain false foreground points that overlap with the projection of the poultry body surface contour, but whose three-dimensional spatial location corresponds to the edge of the operating table, fixed structures, or local depth residues. Therefore, this embodiment uses methods such as RANSAC reference plane fitting and Region of Interest (ROI) selection to refine the initial poultry body surface point cloud. Perform spatial consistency verification and subject screening.

[0047] Specifically, RANSAC reference plane fitting (random sampling consistency) was used to fit the original color point cloud belonging to workbench 7. By performing reference plane fitting, the operating table reference plane is obtained. and on the reference plane of the operating table A local coordinate system is established. Then, the initial point cloud of the poultry body surface is calculated. various points in the middle To the reference plane of the control panel normal distance and points Projected coordinates in the local coordinate system .

[0048] The initial poultry body surface point cloud was filtered by Region of Interest (ROI). Perform filtering and retain normal distance. Points located within the effective normal distance range of the poultry's body surface, and whose corresponding projected coordinates are located within the two-dimensional data acquisition area, i.e., according to the formula:

[0049] ;

[0050] Determine candidate subject point cloud .in, Indicates the reference plane of the control panel The effective normal distance range of poultry body surface determined by the height benchmark (used to eliminate outliers that are close to the operating table or significantly higher than the main body of the poultry body surface). This represents the two-dimensional working area calibrated in the local coordinate system based on the effective acquisition boundary of the operating platform (used to define the initial poultry body surface point cloud on the operating platform reference plane). (effective distribution range).

[0051] Candidate subject point clouds were analyzed using DBSCAN density connectivity clustering. Clustering is performed to group candidate subject point clouds according to density reachability relationships. It is divided into several connected point cloud clusters. Due to the candidate main point cloud... After screening, the main body of the poultry's torso forms the largest continuous point cloud cluster in three-dimensional space, while the remaining edge points, fixed structure points, and local pseudo-foreground points of the operating platform 7 form smaller independent point cloud clusters. Therefore, by counting the number of points contained in each connected point cloud cluster, the connected point cloud cluster with the largest number of points is identified as the main body cluster of the poultry's body. And extract all points from this cluster to construct the main point cloud of the poultry's body surface. .

[0052] S2-1-4. Correct the target poultry body surface mask, and process the main point cloud of the poultry body surface through seed constraint neighborhood retrieval and connectivity consistency verification to generate the effective point cloud of the poultry body surface.

[0053] Because the initial poultry body surface point cloud acquisition process primarily relies on the two-dimensional poultry body surface contour for foreground selection, the low-level regions near the operating table reference plane, located at the longitudinal end of the poultry body surface contour and with sparse, elongated point cloud distribution, are susceptible to mask edge omissions, discontinuous depth boundaries, and interference from nearby fixed structures. This results in some effective low-level foot structures not being included in the main poultry body surface point cloud. Therefore, connected component analysis (CCA), morphological closing, and dilation were used sequentially to mask the surface of the target poultry. The target poultry body surface correction mask is obtained by performing connected component analysis (CCA), morphological closing operation, and dilation operation. .

[0054] Extracting main point clouds from poultry body surfaces The middle satisfies the requirement of being close to the reference plane of the operating table. Furthermore, the points located at the longitudinal ends of the poultry's body outline are integrated into a set of seed points at the bottom of the main body. The set of seed points at the bottom of the main body. Reference plane of the control panel Based on the minimum bounded area covered by the projection point set, the search area for the lower structure is obtained by expanding along the longitudinal end and left and right sides of the poultry body surface. .

[0055] The original color point cloud was analyzed using KD-Tree spatial indexing and K-Nearest Neighbor (KNN) retrieval. Filter and select the low-position height range ( ), planar projection range ( Seed proximity constraint () ) and correction mask constraint (the projected pixel position of a certain point falls into the correction mask of the target poultry body surface) Candidate points (within) are determined according to the formula:

[0056] ;

[0057] Generate low-order candidate point set .in, Represents the original colored point cloud The j-th point To the reference plane of the control panel normal distance, This indicates the allowed height range for lower-level structures. Represents the original colored point cloud The j-th point The plane projection coordinates, Represents the j-th point Seed point set at the bottom of the main body Spatial proximity, Indicates the nearest search radius. Point Projected pixel position Correction mask for falling into the target poultry body surface Inside.

[0058] Target poultry body surface correction mask Masked from the surface of the target poultry Through connected component analysis, morphological closing operations, and dilation operations, its contour boundary is closed and expands outward. Low-position candidate point set. Then from the original color point cloud that has not been subject to selection In, according to height range Low-level structure search area Seed proximity constraints and corrected mask constraints are retrieved again. Therefore, this process, while recalling valid foot points missed due to mask edge omissions or discontinuities in depth boundaries, may also retrieve the set of seed points that coincide with the corrected mask projection and are adjacent to the bottom of the main body. The control panel plate points, fixed structure edge points, and local depth noise points are all included in the low-level candidate point set. .

[0059] For the low-position candidate point set Region growing (RG), spatial connectivity verification, and local planarity discrimination are performed sequentially to remove the platform-like residual points, fixed structure edge residual points, and isolated noise blocks that were included in the candidate point recall process; at the same time, the set of seed points at the bottom of the main body is retained. Effective points that maintain spatial connectivity and conform to the morphological characteristics of the slender structure of the lower foot in local geometric distribution were selected. Finally, the preserved lower structure point cloud was compared with the main point cloud of the poultry body surface. Merging the data yields the effective point cloud of the poultry body surface. .

[0060] S2-2. Based on the effective point cloud of poultry body surface, the scores of spiky external protrusions are calculated by the KNN method (K-nearest neighbor search) and the LPCA method (local principal component analysis). The candidate external protrusions are then subjected to continuous structure verification to suppress spiky external protrusions and generate an optimized poultry body surface point cloud.

[0061] Due to the effective point cloud on the surface of poultry Feather-like protrusions, bristle-like external protrusions, and local depth jitters still exist. A method combining surface bristle suppression and continuous structure verification was adopted to correct local abnormal points that affect subsequent key area segmentation and body size measurement, while preserving the continuous concave-convex features of the keel surface, the thoracic and ventral contour features, and the slender continuous structure of the tarsometatarsus-toe claws.

[0062] S2-2 includes:

[0063] S2-2-1. Construct the local neighborhood of each point in the effective point cloud of poultry body surface using the KNN method, and use the LPCA method to calculate the local normal vector and local curvature of each point to obtain the height residual of each point relative to the local neighborhood body surface.

[0064] It needs to be explained that this is to remove the effective point cloud from the surface of poultry. Feather spikes, bristle-like protrusions, and local depth jitter were analyzed to construct an effective point cloud of poultry body surface using the KNN method. The local neighborhood is then used to calculate the point using the LPCA method. Local normal vector Local curvature Simultaneously, the geometric center of the neighborhood can be obtained from the mean coordinates of the corresponding neighboring points. and along the local normal vector The direction, combined with the geometric center of the neighborhood Calculation points The height residual of the main surface relative to its local neighborhood .

[0065] S2-2-2. Based on the height residual deviation, local curvature deviation, normal consistency and local surface sampling density of each point, calculate the score of the burr-like external protrusion and screen candidate external protrusions according to the score of the burr-like external protrusion.

[0066] According to the formula:

[0067] ;

[0068] ;

[0069] Calculation points Scoring of spiky external protrusions ;in, Represents absolute value. Represents the transpose matrix. , , , All represent weighting coefficients. and Representing points respectively The average height residual and average curvature of each point within the local neighborhood; This represents the local neighborhood obtained through K-nearest neighbor search. Indicates the number of neighboring points. Indicates the first The unit normal vector of each neighboring point This indicates that the normals of the center point and its local neighborhood are consistent. This represents the local surface sampling density, taking the number of K nearest neighbors and the number of points... The distance from the center to the Kth nearest neighbor is the ratio of the area of ​​the local circular region to the radius. The sparser the local point cloud, the better. The smaller; This represents a very small positive number to prevent the denominator from being zero. The lower the normal uniformity or the smaller the local sampling density, the higher the score for the spiky outward protrusion at that point.

[0070] The scores of the spiky external protrusions are sorted from largest to smallest, and the points corresponding to the top M1 scores are selected as candidate external protrusions.

[0071] S2-2-3. Based on the candidate external protrusions, the local principal direction is determined by the LPCA method. Combined with the continuous support on both sides of the local principal direction and the local neighborhood connectivity after temporarily removing the candidate external protrusions, the candidate external protrusions are continuously structurally verified to generate the optimized poultry body surface point cloud.

[0072] For each candidate external protrusion, the corresponding local principal direction is determined using the LPCA method. Along each local principal direction, it is determined whether there are spatially adjacent support points on both sides of the candidate external protrusion with continuous changes in normal and curvature. After temporarily removing these points, the connectivity of the local neighborhood of each candidate external protrusion is checked. Candidate external protrusions lacking continuous support on both sides of the principal direction, and whose removal does not disrupt the connectivity of the local neighborhood, are classified as spiky external protrusions and included in the effective point cloud on the poultry body surface. Spiral-like external protrusions were removed; candidate external protrusions with continuous bilateral support or whose removal would cause local structural breakage were retained, resulting in an optimized poultry body surface point cloud. This improves the boundary preservation capability and local geometric robustness of surface correction under feather interference conditions.

[0073] S2-3. Based on the multilayer perceptron and PointNeXt encoder-decoder backbone network, a DMB-PointNeXt segmentation model is constructed, and the optimized poultry body surface point cloud is processed to generate point cloud data of different regions of poultry.

[0074] The DMB-PointNeXt segmentation model includes a color-geometry dual-branch co-coding module, a scale boundary enhancement feature aggregation encoder, and a decoder.

[0075] The poultry point cloud data for different regions includes point clouds of the keel region, point clouds of the chest-abdomen-wing root composite region, point clouds of the tarsus-toe-claw composite region, and point clouds of other regions.

[0076] If the PointNeXt encoder-decoder backbone network is directly used to process the optimized poultry body surface point cloud, the PointNeXt encoder-decoder backbone network typically inputs spatial coordinates and color information directly in the segmentation of the poultry body surface color point cloud. This leads to problems such as overly strong coupling between color and geometric information, insufficient response to slender keel protrusions, and insufficient ability to handle large curved surfaces in the chest-abdomen-wing root composite region and wing root transition structures, as well as slender small targets in the tarsometatarsal-toe-claw composite region. To solve these problems, this embodiment introduces a color-geometric dual-branch co-coding module, a multi-scale boundary enhancement module, and a joint optimization objective into the PointNeXt encoder-decoder backbone network to perform point-level semantic segmentation of key regions on the poultry body surface.

[0077] In this embodiment, the chest-abdomen-wing root complex region is a continuous surface area within the ventral field of vision of the poultry's body, consisting of the chest region, the abdominal region, and the adjacent regions of the bilateral wing roots, excluding the neck, head, and leg regions. The tarsometatarsal-to-claw complex region is the visible surface area of ​​the distal lower limbs of the poultry's body, extending from the distal end of the tarsal joint to the tips of each toe and claw, including the tarsometatarsal segment, toes, and claws, excluding the femur, the proximal tibia region, and the fixing structure of the operating platform 7.

[0078] S2-3 includes:

[0079] S2-3-1. Construct a color-geometry dual-branch collaborative coding module based on multilayer perceptron and gating fusion strategy;

[0080] Specifically, the color-geometric dual-branch collaborative coding module includes a geometric branch (including a multilayer perceptron), a color branch (including a multilayer perceptron), and a gating fusion strategy.

[0081] The input to the geometric branch is the optimized poultry body surface point cloud. The 3D coordinates, relative height, and local curvature of each point are used to extract the spatial structure features of each point through a multilayer perceptron. The input to the color branch is the optimized poultry body surface point cloud. The RGB information of each point is used to extract the color representation features of the point through an independent multilayer perceptron. Then, an adaptive co-encoding strategy is applied to the spatial structure features and color representation features of each point, enabling the network to automatically adjust the contribution ratio of color information to geometric information based on surface undulations, surface continuity, and color differences in different key regions, resulting in point-level fusion features. The corresponding formula is:

[0082] ;

[0083] ;

[0084] in, This represents the optimized poultry body surface point cloud. The Point-level fusion features of individual points , These represent the optimized poultry body surface point clouds. The Color representation characteristics and spatial structure characteristics of each point This indicates element-wise multiplication. This represents the Sigmoid activation function. Indicates the first Adaptive fusion weights for each point Indicates learnable parameters, This indicates a feature splicing operation.

[0085] S2-3-2. Based on multi-scale grouping (MSG) and edge convolution (EdgeConv), a boundary enhancement feature aggregation module is constructed. Combined with the encoder of the PointNeXt encoder-decoder backbone network, a scale boundary enhancement feature aggregation encoder is generated.

[0086] Specifically, the point-level fusion features of each point output by S2-3-1 are... Compared with the optimized poultry body surface point cloud Corresponding point The three-dimensional coordinates are paired one-to-one according to the point index and used as the first-level input of the PointNeXt encoder. Subsequent coding layers use the point set and point features output by the previous coding layer as the input of the current layer. In the downsampling set abstraction layer of the PointNeXt encoder, the center point is selected using the farthest point sampling (FPS) method, with the current input point set as the support point set.

[0087] Using each center point as the sphere's center, a ball query (BQ) is employed, with the search radius calculated on a small scale. Mesoscale search radius and large-scale search radius By retrieving neighborhood points from the current support point set, local point sets at three scales are obtained, among which... Whether a point belongs to the corresponding local point set is determined by its three-dimensional coordinate distance from the center point, and each neighboring point carries the local input features corresponding to its point index.

[0088] In local point sets at each scale, the relative coordinate difference and feature difference between neighboring points and the center point are calculated. The center point features, relative coordinate differences, and feature differences are input into the shared multilayer perceptron corresponding to each scale, and the EdgeConv graph edge relationship encoding is completed through max pooling to obtain small-scale local features. Local features at the mesoscale and large-scale local features By jointly representing multi-scale spatial relationships and local feature differences, the ability to discriminate the slender protruding boundaries of the keel, the continuous curved surface contours of the thorax-wing root complex region, and the slender structures of the tarsometatarsus-toe claw complex region is enhanced. The three scale features are adjusted to the same channel dimension through point-by-point linear mapping, and then adaptively fused using scale attention.

[0089] ;

[0090] ;

[0091] in, Indicates the first Multi-scale enhancement features at each center point , , These represent the adaptive weights for small-scale, medium-scale, and large-scale operations, respectively (the sum of the three is 1). Represents the normalized exponential function, Represents the scale weight learning matrix. This indicates feature splicing.

[0092] S2-3-3, a DMB-PointNeXt segmentation model is constructed by jointly optimizing an objective based on a color-geometry dual-branch co-coding module, a scale boundary enhancement feature aggregation encoder, a decoder, and a segmentation head;

[0093] An initial DMB-PointNeXt segmentation model is constructed based on a color-geometric dual-branch co-coding module, a scale boundary enhancement feature aggregation encoder, a decoder, and a segmentation head. The initial DMB-PointNeXt segmentation model is then trained using a joint optimization objective to obtain the final DMB-PointNeXt segmentation model.

[0094] Considering that the keel region and the tarsometatarsus-toe claw composite region account for a small proportion of the overall point cloud, and that the keel boundary is thin and has obvious local undulations, the joint optimization objective is... Including weighted cross-entropy loss (WCE) Dice loss Boundary loss (BL) To simultaneously enhance small-sample class recognition ability, overall segmentation integrity of key regions, boundary prediction stability, and training robustness under class imbalance conditions, the corresponding formula is:

[0095] ;

[0096] in, , These represent the Dice loss weights and the boundary loss weights, respectively.

[0097] The multi-scale enhanced features of each point are input into the DMB-PointNeXt segmentation model to generate point clouds of the keel region. Point cloud of the thoracic-abdomen-pterygoid root complex region Point cloud of tarsometatarsal-to-claw complex region Point clouds in other regions.

[0098] S3. Based on point cloud data of different regions of poultry, calculate the keel length parameter, chest body size parameter and tibia length; the keel length parameter includes the centerline arc length and straight keel length; the chest body size parameter includes the anterior chest width, posterior chest width and chest length;

[0099] S3 includes:

[0100] S3-1. Based on the point cloud of the keel region, construct the centerline of the poultry keel. Combine the front-end constraint and the back-end steep slope response score to locate the front-end point and the rear-end point respectively, and obtain the straight keel length and the centerline arc length and keel length.

[0101] S3-1 includes:

[0102] S3-1-1. Based on the point cloud of the keel region, PCA (Principal Component Analysis) is used to determine the longitudinal main direction and centerline nodes of the keel. The centerline nodes are connected in longitudinal order to obtain the poultry keel centerline.

[0103] Specifically, the point cloud of the keel region is analyzed on the YZ plane using PCA, the corresponding covariance matrix is ​​calculated and eigenvalue decomposition is performed, and the eigenvector corresponding to the largest eigenvalue is taken as the longitudinal principal direction of the keel. The point cloud of the keel region is then projected along the longitudinal principal direction of the keel, according to the formula:

[0104] ;

[0105] Generate the first The first keel point (the first point in the keel region point cloud) Projected coordinates of each point along the longitudinal main direction of the keel ;in, This indicates the mean center of the keel point in the YZ plane. Indicates the longitudinal main direction of the keel. , They represent the first Y-axis and Z-axis coordinates of each keel point This represents the dot product of vectors.

[0106] Based on the projected coordinates of each keel point along the longitudinal main direction of the keel, the disordered point cloud of the keel region is arranged and transformed into an ordered set of points arranged along the keel extension direction. This is done according to the projected coordinates. The distribution of the point cloud in the keel region was used to perform equidistant slices, resulting in multiple continuous longitudinal cross-sectional point sets. Continuous longitudinal cross-sectional point sets with fewer points than the threshold or abnormal distribution are removed to determine the effective cross-sections; the three-dimensional centroid is calculated for each effective cross-section. And select the actual keel point closest to the centroid from the corresponding effective cross section as the centerline node. The corresponding formula is:

[0107] ;

[0108] in, Indicates the first Centerline nodes corresponding to each effective cross section Represents a continuous longitudinal section point set The three-dimensional coordinates of the j-th actual keel point. Represents Euclidean distance. This indicates the operation of taking the minimum value.

[0109] Connect the centerline nodes in a longitudinal order to obtain the poultry keel centerline. This ensures that the centerline extends continuously along the main extension direction of the keel, and that the centerline nodes always fall on the actual keel point cloud, thereby improving the geometric consistency and measurement repeatability of the centerline construction.

[0110] S3-1-2, Based on the poultry keel centerline and keel region point cloud, through The shape boundary (Alpha Shape) is extracted to obtain the leading edge boundary point cluster, and the leading edge point is determined by combining the leading edge constraint;

[0111] Specifically, in this embodiment, under a fixed acquisition posture, the side with the smaller Y-coordinate value of the node in the center line of the poultry keel is defined as the front end of the keel, and the side with the larger Y-coordinate value is defined as the rear end of the keel.

[0112] Based on the front end of the keel, candidate regions for the front end of the point cloud of the keel region are extracted according to the Y-direction range, and then the candidate regions are processed in the XY plane. Shape boundary extraction to obtain leading edge boundary point clusters Then, take the first node of the center line of the poultry keel. For reference, the actual keel points that satisfy the front-end constraints (adjacency relationship between the forward boundary position and the centerline) are selected as the front-end points. This ensures that the front point is located at the visible leading edge of the keel while maintaining structural consistency with the front end of the centerline. The corresponding formula is:

[0113] ;

[0114] in, , These represent the weight coefficients of the forward boundary term and the centerline proximity term, respectively. Indicates the first candidate in the front-end candidate area The pixel positions of the candidate boundary points Indicates the first The forward and backward coordinates of the candidate boundary points.

[0115] S3-1-3. Based on the centerline of the poultry keel, calculate the Z-direction change rate, slope angle response and local curvature response of adjacent nodes, construct the steep slope response score at the rear end and determine the rear endpoint;

[0116] Specifically, the Z-axis rate of change between adjacent nodes within the tail segment of the poultry keel centerline is calculated. . Node With nodes The absolute value of the Z-coordinate difference divided by twice the equidistant slice interval yields the Z-axis rate of change. Then the slope angle response was calculated. and local curvature response Then, according to the formula:

[0117] ;

[0118] = ;

[0119] = ;

[0120] ;

[0121] Calculate the first Back slope response score of each centerline node ;in, Represents the arctangent function. These represent the Z-axis weight, slope angle weight, and local curvature weight, respectively. They represent the first Each centerline node The x and y coordinates in the plane They represent the first The, the The and the first Each centerline node Projected coordinates in the plane; This indicates the slicing interval used when performing equidistant slicing along the longitudinal projection coordinates of the keel in S3-1-1; The larger the value, the steeper the slope; The larger the value, the more obvious the local bending of the center line.

[0122] Savitzky-Golay (SG) smoothing was applied to the response scores of each steep slope at the back end to suppress local fluctuation noise and preserve the main peak shape, resulting in a smoothed response sequence. For smooth response sequences Non-maximum suppression (NMS) is performed to remove adjacent duplicate peaks and retain the local dominant response peak. The centerline node with the largest response value retained by NMS is used as the back-end reference point. Then, radius search and nearest neighbor selection are used to extract back-end candidate clusters in its local neighborhood. From these, the actual keel point that is structurally continuous with the reference point and extends along the keel extension direction is determined as the back-end endpoint. This improves the stability, anti-interference capabilities, and peak value determination robustness of the backend positioning system.

[0123] S3-1-4. Based on the centerline, front end point, and rear end point of the poultry keel, calculate the straight keel length and the centerline arc length of the keel.

[0124] Specifically, the three-dimensional Euclidean distance between the front end point and the rear end point is denoted as the straight keel length. .

[0125] According to the formula:

[0126] ;

[0127] Calculate the centerline arc length and keel length ;in, These represent the centerline nodes corresponding to the front and rear endpoints, respectively. This represents the summation function. They represent the first The centerline node and the first One centerline node This indicates the path length between adjacent nodes on the centerline.

[0128] S3-2. Based on the point cloud of the thorax-abdomen-wing root composite region and the posterior endpoint, the anterior chest width, posterior chest width and chest length are determined by the positioning of anterior chest feature points, the positioning of posterior edge local corners and the constraint of anterior edge baseline.

[0129] S3-2 includes:

[0130] S3-2-1. Based on the point cloud of the chest-abdomen-wing root composite region, the semantic candidate sets of the left anterior chest and the right anterior chest are determined and the vertical distance is calculated by smoothing through the USIP network (unsupervised stable interest point detection network) and local weighted regression (LOESS).

[0131] Due to the differences in local morphology among different poultry individuals, and the fact that the left and right measurement points are not simply symmetrical extreme points, the point cloud of the thoracic-abdomen-wing root complex region... The search area is selected by intercepting the area located on the front side of the poultry's body surface and near the left and right wing roots. The projection of the poultry keel centerline onto the XY plane is used as the axis of the poultry body surface. .

[0132] front edge search domain As input to the USIP network, a set of candidate keypoints for the leading edge is generated. and the key point uncertainty corresponding to each candidate key point of the leading edge. .in, The smaller the value, the higher the localization stability of the leading edge candidate keypoints under point cloud perturbation and resampling conditions. The USIP network is an existing network structure, so it will not be described in detail.

[0133] For the candidate keypoint set of the leading edge Perform mutual reconstruction constraints (MR) to obtain the first... Self-reconstruction error of candidate key points at the leading edge and cross-sample mutual reconstruction error Among them, the self-reconstruction error is used to characterize the positional consistency of the leading candidate keypoint before and after reconstruction in the same sample, and the cross-sample mutual reconstruction error is used to characterize the semantic correspondence stability of the leading candidate keypoint between different poultry individuals.

[0134] Based on the leading edge candidate key points being located on the central axis of the poultry body surface. The left and right sides are used to set the candidate keypoints of the leading edge. Divided into left anterior chest semantic candidate set and right anterior chest semantic candidate set .

[0135] Extracting the semantic candidate set of the left anterior chest and right anterior chest semantic candidate set The nearest anterior boundary points are identified, and local weighted regression (LOESS) is performed in the XY plane to smooth the boundary curves, constructing locally continuous boundary curves for the left and right anterior edges. Then, a planar Frenet frame is used to calculate the boundary curves within the semantic candidate set of the left anterior chest. and right anterior chest semantic candidate set Local curvature at the corresponding positions of each candidate point And calculate the axis from the candidate points to the poultry body surface. vertical distance .

[0136] S3-2-2. Calculate the comprehensive response function of the anterior chest measurement points of each candidate point in the left anterior chest semantic candidate set and the right anterior chest semantic candidate set respectively, and determine the feature points of the left anterior chest and the right anterior chest.

[0137] Specifically, the comprehensive response function of the anterior chest measuring point The corresponding formula is:

[0138] ;

[0139] ;

[0140] ;

[0141] ;

[0142] ;

[0143] ;

[0144] in, These represent the non-negative weights corresponding to the stability response, semantic consistency response, curvature response, and lateral extension response, respectively (the sum of the four weights is 1). Representing the semantic candidate sets of the left anterior chest and the right anterior chest respectively, the first... The stability response of each candidate point, the semantic consistency response (the smaller the sum of the self-reconstruction error and the cross-sample mutual reconstruction error, the larger the response), and the lateral abduction response (the larger the vertical distance from the candidate point to the axis of the poultry body surface, the larger the response). Indicates the first The curvature response of each candidate point at the leading edge boundary location (the more pronounced the local boundary transition, the larger the response). Represents the semantic candidate set of the left anterior chest. Or right anterior chest semantic candidate set , This represents the normalization operator (which eliminates dimensional differences between keypoint uncertainty, mutual reconstruction error, local curvature, and lateral distance). All of these represent parameters to be normalized.

[0145] semantic candidate set of the left anterior chest Selected from The candidate point with the largest value is selected as the left anterior chest feature point. In the semantic candidate set of the right anterior chest Selected from The candidate point with the largest value is selected as the right anterior chest feature point. .

[0146] S3-2-3, Calculate the Euclidean distance between the left and right anterior chest feature points and use it as the anterior chest width;

[0147] S3-2-4. Based on the point cloud of the thorax-abdomen-wing root composite region, the corner points connecting the left leg and thorax-abdomen and the right leg and thorax-abdomen are determined by cubic B-spline fitting (CBS) and local tangent direction analysis.

[0148] Specifically, point clouds were generated from the thoracic-abdomen-pterygoid root complex region. Extract the set of posterior boundary points on the poultry body surface and project it onto... The plane is used to obtain the set of trailing edge contour points. The central axis of the poultry body surface is determined based on S3-2-1. The trailing edge profile is divided into a left trailing edge point set and a right trailing edge point set. A continuous trailing edge profile curve is constructed using cubic B-spline fitting (CBS) for each set. The curve traversal direction is then standardized from the front to the rear of the poultry, and arc length parameterization (AP) is performed to obtain the left trailing edge parameter curve. And right trailing edge parameter curve .

[0149] right and Perform local rotation analysis separately at each effective arc length position. The distance between the two sides is a preset arc length of and Calculate the first derivative of the curve at each point and normalize it to obtain the corresponding local unit tangent vector. and And calculate the arc length position according to the following formula. Local angle value at :

[0150] ;

[0151] Where, when the angle between two local unit tangent vectors is relatively small... The larger the value, the more obvious the local turning point at that location; This represents the inverse cosine function.

[0152] For any candidate point, calculate the axis of that candidate point in the poultry body surface. The nearest projection point on the candidate point, and the unit vector pointing from the candidate point to the nearest projection point is denoted as . ;when At that time, the candidate point was determined as a concave corner candidate point turning towards the central axis of the poultry body surface. Selected on the left and right trailing edge parameter curves respectively... The candidate point with the largest concave corner value is determined. Left leg-chest-abdomen connection corner point in the plane And the corner point connecting the right leg and chest / abdomen .

[0153] S3-2-5. Calculate the Euclidean distance in the XY plane between the left leg-chest-abdomen connection point and the right leg-chest-abdomen connection point, and use it as the final chest width. In this embodiment, S3-2-4 and S3-2-5 utilize the bilateral concave corner structure formed at the connection between the posterior edge of the thorax-abdomen-wing root complex region and the legs. Through continuous expression of the posterior edge contour, analysis of local tangent direction, and positioning of the bilateral concave corner, adaptive determination of the bilateral measurement points of the final chest width is achieved, thereby improving the geometric consistency and repeatability of the final chest width measurement.

[0154] S3-2-6. Based on the left anterior chest feature point, the right anterior chest feature point, and the posterior endpoint, construct the anterior baseline and calculate the chest length using orthogonal projection (OP).

[0155] Specifically, in the measurement of breast length in poultry, the anterior boundary is not a single discrete point, while the posterior boundary can be determined from the posterior endpoint. To ensure stable anchoring and that chest length measurements reflect the longitudinal span from the anterior boundary to the posterior end of the chest, feature points on the left anterior chest were extracted. Right anterior chest feature points The local boundary points of the adjacent leading edge are merged into a set of leading edge points. Subsequently, regarding Perform total least squares fitting (TLS) to construct the leading edge baseline. Then... Towards Perform orthographic projection (OP) to obtain the projection points. .because It can stably characterize the posterior position of the caudal thorax, therefore... As the anchor point on the side of the thorax and tail, and with As an expression of the anterior boundary of the chest length, it enables chest length measurement constrained by both the anterior boundary and a single point at the posterior end, and... To its Upper projection point The orthogonal distance between them is denoted as chest length. This embodiment achieves stable measurement of poultry chest length by fitting the leading edge baseline, orthogonally projecting the rear anchor point, and constraining the chest major axis, and improves the robustness of chest length measurement to local fluctuations at the leading edge and single-point positioning errors at the rear.

[0156] S3-3. Based on the point cloud of the tarsometatarsal-to-claw composite region, the starting point of tibia length is located by the CEDT method and the CCS method, and the ending point of tibia length is located by the QSF method and SG smoothing method, and the corresponding tibia length is calculated.

[0157] S3-3 includes:

[0158] S3-3-1. Based on the point cloud of the tarsometatarsal-to-claw composite region, the main branches of the tarsometatarsal trunk and the forward toe branches are determined. Combining the CEDT method (constrained Euclidean distance transformation method) and the CCS method (constrained circular kernel search method), a candidate circular kernel hierarchical screening strategy with coverage priority is adopted to determine the tibia length starting point.

[0159] Specifically, considering that the tibia length origin is not a simple boundary extreme point, but is located in the metatarsophalangeal transition zone and falls in the direction of the middle toe corresponding to the extension line of the tarsometatarsal main axis, the point cloud of the tarsometatarsal-to-claw composite region is... Projecting onto the XY plane yields the set of projection points for the feet. The shape boundary extraction (Alpha Shape) constructs the outer contour of the foot from the set of foot projection points, and the corresponding foot planar skeleton is obtained through the median transformation (MAT).

[0160] Topological branching analysis was performed on the foot's planar skeleton to identify the tarsometatarsal main branches and anterior digitigrade branches. Based on the cumulative length and extension direction of each branch in the foot's planar skeleton, the longest branch extending from the metacarpophalangeal junction towards the proximal tarsometatarsal segment was identified as the tarsometatarsal main branch, and the branches extending from the metacarpophalangeal junction towards the tips of each toe were identified as anterior digitigrade branches. The terminal skeletal point of the tarsometatarsal main branch before entering the metacarpophalangeal junction was identified as the tarsometatarsal main branch tip, and the starting skeletal point of each toe branch after leaving the metacarpophalangeal junction was identified as the corresponding toe base skeletal positioning point. The metacarpophalangeal transition zone was defined by the tarsometatarsal main branch tip and the toe base skeletal positioning points. .

[0161] Least squares fitting (TLS) was performed on the tarsometatarsal main branches to obtain the tarsometatarsal principal axis. Then Extending forward to the metatarsophalangeal transition zone Considering that there are four toe branches at the front of the foot, the toe number is not used as the criterion. Instead, the branch with the smallest angle to the tarsometatarsal axis, the smallest lateral offset, and the strongest continuity with the tarsometatarsal trunk is selected from the forward toe branches as the corresponding branch of the middle toe, thus obtaining the extension line of the middle toe. .

[0162] Using the skeletal positioning points at the base of each toe as location references, corresponding local boundary segments are extracted from the ordered foot outline generated by Alpha Shape to obtain the connection regions at the base of each toe; simultaneously, the metatarsophalangeal transition area is extracted. The orderly boundary points located on one side of the palm form the palm-to-toe transition arc.

[0163] The base of the fourth toe connects to the palm. Inside, only along A candidate circle center sequence is constructed, and the adaptive radius corresponding to each candidate circle center is calculated by combining the constrained Euclidean distance transformation (CEDT). Then, constrained circle kernel search (CCS) is used to perform hierarchical evaluation on each candidate circle kernel.

[0164] According to the formula:

[0165] ;

[0166] Calculate the first Toe base coverage response of candidate circular nuclei .

[0167] According to the formula:

[0168] ;

[0169] Calculate the first Palm-to-toe transition arc fitting response of candidate circular nuclei .

[0170] According to the formula:

[0171] ;

[0172] Calculate the first Proximal position response of candidate circular nuclei .in, Indicates the first The number of toe base connection regions covered by the circumferential neighborhood of each candidate nucleus; This indicates the total number of toe base connection areas identified, corresponding to the complete foot. =4; Indicates the first The adaptive radius of each candidate circular nucleus; Indicates the distance from each boundary point of the palm-toe transition arc to the first... The average absolute radial deviation of each candidate circumference; Indicates the first The distance between the center of each candidate circle and the end of the tarsal trunk; Indicates the extension line of the middle toe In the metatarsophalangeal transition zone The length of the candidate circle center search interval within the range. The larger the values ​​of the three responses, the better the corresponding evaluation result.

[0173] First, keep The largest candidate circular nucleus; then select from it. The largest candidate circular kernel; when multiple identical results exist, select the largest one. The largest candidate nucleus was identified, and the center of the final candidate nucleus was determined as the tibia length starting point. .

[0174] S3-3-2. Based on the point cloud of the tarsometatarsal-to-claw composite region, the tibia length endpoint is generated by using the QSF method (local quadratic surface fitting method) and principal curvature estimation combined with SG smoothing.

[0175] The tibia length endpoint is located in the proximal bony prominence of the tarsometatarsal segment, corresponding to the highest outward position of the local bony prominence. Extracting from the location The point cloud of the tarsometatarsal segment on the proximal side was obtained, and an equally spaced orthogonal slice (OS) sequence was established along the tarsometatarsal principal axis. The first... Project the points within each slice onto the corresponding orthogonal slice plane, and calculate the two-dimensional convex hull area of ​​the projected point set. And the average two-dimensional convex hull area of ​​several adjacent slices on the distal side of the slice. As a local slender trunk reference, and then according to the formula:

[0176] ;

[0177] Calculate the first The degree of cross-sectional expansion of each slice , The larger the value, the more pronounced the cross-sectional expansion at that location. Indicates the first The area of ​​the two-dimensional convex hull of each slice; Indicates the first The average two-dimensional convex hull area of ​​several adjacent slices on the far side of a slice.

[0178] A cross-sectional expansion sequence was constructed according to the arrangement order of each slice on the tarsometatarsal principal axis. Savitzky-Golay smoothing (SG smoothing) was performed on the cross-sectional expansion sequence to reduce the influence of local point cloud fluctuations on the cross-sectional area calculation. Then, non-maximum suppression (NMS) was used to retain the main response peak of the proximal region, and the slice corresponding to the main response peak and its adjacent slices were identified as the proximal bony prominence region.

[0179] Local quadratic surface fitting (QSF) was used to reconstruct the local surface within the proximal bony prominence region, and the two principal curvatures of the fitted surface at each location were calculated downwards using a uniform outward method. The formula is as follows:

[0180] ;

[0181] Calculate the position of the fitted surface Surface protrusion strength .in, Represents the maximum value function; and Representing the fitted surface positions respectively The two principal curvatures, the average of which represents the average curvature at that location, are used to preserve the outward convex response; The larger the value, the more pronounced the bony protrusion at that location.

[0182] exist Within the convex surface region where >0, select the point with the maximum outward height of the QSF fitted surface and determine it as the tibia length endpoint. .

[0183] S3-3-3 Calculate the three-dimensional Euclidean distance between the tibia length start point and the tibia length end point, and use it as the tibia length. .

[0184] Step S3-3 uses the center of the round nucleus in the palmar-toe transition zone as the distal measurement reference and the center of the proximal bony prominence of the tarsometatarsal segment as the proximal measurement reference. This avoids the problem of measurement point drift caused by relying solely on boundary extreme points or manual experience in point selection, thereby improving the stability and repeatability of poultry shank length measurement.

[0185] S4. Based on basic physical characteristics data, keel length parameters, chest body size parameters, and tibia length, calculate the corresponding comprehensive development score, and generate poultry management results by combining the GMM model and DS evidence theory.

[0186] S4 includes:

[0187] S4-1. Construct a decision feature vector based on basic vital signs data, keel length parameters, chest body size parameters, and tibia length.

[0188] Specifically, using individual identification tags (RFID numbers) as poultry individual identification indexes, weight, rectal temperature, keel length, chest body size, and shank length data collected from the same poultry individual within the same testing batch are timestamped and linked to establish individual phenotypic records. These individual phenotypic records undergo standardized unit verification, outlier checks, and batch-specific statistical processing. The median rectal temperature of poultry individuals within the same testing batch is used as the reference rectal temperature. Calculate the current rectal temperature of individual poultry. Compared with reference rectal temperature The absolute difference between them is recorded as the rectal temperature deviation. .

[0189] Calculate the ratios of the centerline arc length, keel length, straight keel length, forechest width, end-chest width, chest length, and tibia length to the body weight data, and integrate all the ratios to form the current poultry decision feature vector;

[0190] S4-2. Based on the decision feature vector, calculate the standardized feature values, combine subjective weights and objective weights to determine the set of key volume scale anomalies and calculate the weighted Euclidean distance;

[0191] Specifically, the decision feature vector is standardized using Z-score to obtain the standardized feature value of the j-th body size-to-weight ratio for the i-th poultry individual. The deviation of the above six standardized characteristic values ​​(standardized characteristic values ​​of centerline arc length, keel length, straight keel length, anterior chest width, distal chest width, chest length, and tibia length) is checked, and the absolute standardized characteristic values ​​are determined. The body size to body weight ratio indicators greater than 2 (the ratios of centerline arc length, keel length, straight keel length, forechest width, endchest width, chest length, and tibia length to body weight data, respectively) are taken as the key body size anomalies of the i-th poultry individual, and the corresponding anomaly indicator name, deviation direction, and absolute standardized deviation amount are combined to form a set of key body size anomalies.

[0192] Based on the level of attention paid to keel development, chest development, and tarsometatarsal development in breeding or production management, subjective weights were assigned to the above six body size-to-weight ratios. The sum of the six subjective weights is set to 1. For the j-th item, body size to weight ratio, the original ratio before Z-score standardization is used to calculate its mean and standard deviation among poultry individuals in the same batch. The ratio of the standard deviation to the mean is used as the coefficient of variation for this item to characterize the degree of information dispersion. The coefficient of variation of this item is divided by the sum of the coefficients of variation of the six items to obtain the objective weight. The larger the coefficient of variation, the greater the objective weight of that indicator.

[0193] According to the formula:

[0194] ;

[0195] ;

[0196] ;

[0197] Calculate the first The weighted Euclidean distance between each poultry and the positive / negative ideal solution .in, This indicates the proportion of subjective weight ( ), This represents the comprehensive weight of the j-th decision feature in the decision feature vector. This represents the decision feature value after applying the comprehensive weights. This represents the positive / negative ideal value of the j-th decision feature (the maximum and minimum values ​​of the corresponding columns of the weighted decision matrix are taken after the index is homogenized). This represents the total number of decision features. A weighted decision matrix refers to using the m poultry individuals from the same batch as rows of a matrix, and... (In this embodiment) =6) decision features are used as matrix columns, and each Arranged according to the corresponding individual index and feature index.

[0198] S4-3. Based on weighted Euclidean distance, the comprehensive developmental score is calculated by improving the TOPSIS comprehensive evaluation method.

[0199] The formula corresponding to the improved TOPSIS comprehensive evaluation method is:

[0200] ;

[0201] in, , They represent the first A single poultry and the ideal solution, the first The weighted Euclidean distance between each poultry and the negative ideal solution. Indicates the first The overall development score of each poultry.

[0202] The closer the overall development score is to 1, the closer the individual poultry is to the excellent development status of the same batch in terms of weight, keel development, chest development, and tarsometatarsal development.

[0203] S4-4. Based on the comprehensive developmental score and the set of key body size abnormalities, poultry management results are generated through the GMM model and DS evidence theory.

[0204] To avoid inconsistencies in individual evaluation standards for poultry of different batches and ages due to relying solely on fixed empirical thresholds, the comprehensive developmental score is input into a Gaussian Mixture Model (GMM) for clustering. The GMM sets four mixture components and estimates the mean, variance, and mixture weight of each component based on the distribution characteristics of the comprehensive developmental scores within the same batch, outputting the first... The posterior probability that each poultry species belongs to each of the mixture components.

[0205] The four mixed components were sorted from highest to lowest according to the mean of the comprehensive developmental score, corresponding to excellent development, normal development, sub-healthy development, and abnormal developmental status, respectively. The fourth component... The posterior probabilities of each poultry belonging to one of the above four states form the state posterior probability vector, and the state with the highest posterior probability is taken as the initial state level of the poultry individual.

[0206] Using the posterior probability vector of the state as evidence of individual developmental state, and the corresponding anal temperature deviation... As evidence of body temperature status, a set of key body size abnormalities is used as evidence of body size abnormalities. These are then fused using the DS evidence theory to obtain basic probability allocation values ​​for four states: excellent development, normal, sub-healthy, and abnormal concern. These basic probability allocation values ​​together constitute the DS fusion result. The state corresponding to the highest basic probability allocation value is determined as the final individual state judgment result, and this highest basic probability allocation value is determined as the decision confidence level. The lower limit of the confidence level is determined based on the verification samples that have completed artificial state confirmation; in this embodiment, it is set to 0.7.

[0207] When the final individual status assessment result is excellent development and the decision confidence level is not lower than 0.70, a poultry management result of "breeding candidate or key breeding stock" is generated; when the final individual status assessment result is normal and the decision confidence level is not lower than 0.70, a poultry management result of "maintain current feeding management" is generated; when the final individual status assessment result is sub-healthy or abnormal concern and the decision confidence level is not lower than 0.70, a poultry management result of "key retesting, group feeding, health check or culling concern" is generated; when the decision confidence level is lower than 0.70, this status assessment is marked as a low confidence result, and a prompt for retesting and manual verification is generated.

[0208] In summary, this invention constructs a data acquisition device integrating RFID identification, weight-temperature detection, depth camera, and industrial all-in-one machine. This achieves unified acquisition and numbering of individual poultry identity, basic physical characteristics, and visual data within the same measurement process, avoiding manual recording and data mismatch issues. Through RGB mask guidance, spatial consistency verification, and structure preservation correction, it solves problems such as live poultry point cloud mixing with the background and interference from feather protrusions, improving the extraction quality and continuity of effective point clouds from the body surface. Utilizing the DMB-PointNeXt segmentation model, it achieves precise point-level semantic segmentation of the keel, chest-abdomen-wing root composite region, and tarsometatarsus-toe composite region. Combined with customized body size extraction methods such as keel centerline construction and anterior / posterior endpoint positioning, key feature point positioning of anterior and posterior chest width, and tibia length measurement constrained by palm-to-toe transition and proximal bony prominences, it realizes automated and standardized measurement of core body size parameters such as keel length, chest length, chest width, and tibia length.

[0209] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0210] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multimodal sensing-based intelligent measurement and management method for poultry body size and weight, characterized in that, include: A multimodal phenotypic data collection device for poultry was set up, and basic vital signs data and RGB point cloud data of poultry were collected. The RGB point cloud data includes RGB images, depth images, and raw color point cloud data; Based on RGB point cloud data, seed-constrained neighborhood retrieval and connectivity consistency verification are used to segment the data using the DMB-PointNeXt segmentation model, generating point cloud data for different regions of poultry. The point cloud data for different regions of poultry includes point clouds of the keel region, point clouds of the chest and abdomen-wing root composite region, point clouds of the tarsus-toe and claw composite region, and point clouds of other regions. Based on point cloud data of different regions of poultry, the keel length parameter, chest body size parameter, and tibia length are calculated; the keel length parameter includes the centerline arc length and straight keel length; the chest body size parameter includes the anterior chest width, posterior chest width, and chest length. Based on basic physical characteristics data, keel length parameters, chest body size parameters, and tibia length, the corresponding comprehensive development score is calculated, and poultry management results are generated by combining the GMM model and DS evidence theory.

2. The intelligent measurement and management method for poultry body size and weight based on multimodal sensing according to claim 1, characterized in that, The poultry multimodal phenotypic acquisition equipment includes an inlet RFID reading module (1), a weight weighing platform (2), a temperature detection unit (3), a depth camera (4), an industrial all-in-one machine (5), a mobile power supply (6), an operating table (7), a vertical support (8), and an LED supplementary light (9).

3. The intelligent measurement and management method for poultry body size and weight based on multimodal sensing according to claim 1, characterized in that, The generation of point cloud data for different regions of poultry includes: The RGB point cloud data was processed using the Mask2Former model and the DBSCAN density connectivity clustering method. Combined with seed constraint neighborhood retrieval and connectivity consistency verification, effective point clouds of poultry body surface were generated. Based on the effective point cloud of poultry body surface, the scores of spiky external protrusions are calculated by the KNN method and LPCA method, and the candidate external protrusions are continuously verified to generate an optimized poultry body surface point cloud. Based on a multilayer perceptron and a PointNeXt encoder-decoder backbone network, a DMB-PointNeXt segmentation model is constructed, and the optimized poultry body surface point cloud is processed to generate point cloud data for different regions of poultry.

4. The intelligent measurement and management method for poultry body size and weight based on multimodal sensing according to claim 3, characterized in that, The generation of the optimized poultry body surface point cloud includes: The local neighborhood of each point in the effective point cloud of poultry body surface is constructed by KNN method, and the local normal vector and local curvature of each point are calculated by LPCA method to obtain the height residual of each point relative to the body surface of the local neighborhood. Based on the height residual deviation, local curvature deviation, normal consistency and local surface sampling density of each point, the score of the burr-like external protrusion is calculated, and candidate external protrusions are screened according to the score of the burr-like external protrusion. Based on candidate external protrusions, the local principal direction is determined by LPCA analysis. Combined with the continuous support on both sides of the local principal direction and the local neighborhood connectivity after temporarily removing candidate external protrusions, the candidate external protrusions are continuously structurally verified to generate an optimized poultry body surface point cloud.

5. The intelligent measurement and management method for poultry body size and weight based on multimodal sensing according to claim 3, characterized in that, The construction of the DMB-PointNeXt segmentation model includes: Based on a multilayer perceptron and a gating fusion strategy, a color-geometry dual-branch collaborative coding module is constructed. Based on multi-scale grouping and edge convolution, a boundary enhancement feature aggregation module is constructed, which is then combined with an encoder to generate a scale boundary enhancement feature aggregation encoder. A DMB-PointNeXt segmentation model is constructed based on a color-geometric dual-branch co-coding module, a scale boundary enhancement feature aggregation encoder, a decoder, and a segmentation head, through a joint optimization objective.

6. The intelligent measurement and management method for poultry body size and weight based on multimodal sensing according to claim 1, characterized in that, The calculation of keel length parameters, chest body size parameters, and tibia length includes: Based on the point cloud of the keel region, the centerline of the poultry keel is constructed. The front point and the rear point are located by combining the front constraint and the rear steep slope response score, respectively, and the straight keel length and the centerline arc length and keel length are obtained. Based on the point cloud of the thorax-abdomen-wing root composite region and the posterior endpoint, the anterior chest width, posterior chest width and chest length are determined by the localization of anterior chest feature points, the local corner localization of the posterior edge and the constraint of the anterior baseline. Based on the point cloud of the tarsometatarsal-to-claw composite region, the tibia length starting point is located by the CEDT and CCS methods, and the tibia length ending point is located by the QSF and SG smoothing methods, and the corresponding tibia length is calculated.

7. The intelligent measurement and management method for poultry body size and weight based on multimodal sensing according to claim 6, characterized in that, The obtained straight keel length and centerline arc length keel length include: Based on the point cloud of the keel region, PCA is used to determine the longitudinal main direction and centerline nodes of the keel. The centerline nodes are connected in longitudinal order to obtain the poultry keel centerline. Based on the poultry keel centerline and keel region point cloud, through Shape boundary extraction yields a cluster of leading edge boundary points, which are then combined with leading edge constraints to determine the leading edge points. Based on the centerline of the poultry keel, the Z-direction change rate, slope angle response and local curvature response of adjacent nodes are calculated, the steep slope response score of the rear end is constructed and the rear endpoint is determined; Calculate the straight keel length and the centerline arc length of the keel based on the centerline, front end point, and rear end point of the poultry keel.

8. The intelligent measurement and management method for poultry body size and weight based on multimodal sensing according to claim 1, characterized in that, The generated poultry management results include: Based on basic vital signs data, keel length parameters, chest body size parameters, and tibia length, a decision feature vector is constructed. Based on the decision feature vector, standardized feature values ​​are calculated, and subjective and objective weights are combined to determine the set of key volume scale anomalies and calculate the weighted Euclidean distance. Based on weighted Euclidean distance, an improved TOPSIS comprehensive evaluation method is used to calculate the comprehensive developmental score; Based on the comprehensive developmental score and the set of key body size abnormalities, poultry management results are generated using the GMM model and DS evidence theory.

Images