Intelligent identification and grouping guidance system suitable for camel milking parlour

Abstract

The invention discloses an intelligent identification and grouping guidance system suitable for a camel milking parlour, and belongs to the technical field of grouping guidance, and the system comprises a data collection module which is used for obtaining visual feature data and identification data of camels; the feature processing module is used for analyzing the identification data to generate candidate identity numbers and generate an initial identity data set; the verification module is used for starting a hierarchical verification process to screen and exclude candidate identities in the initial identity data set until only unique candidate identities are left in the initial identity data set; the grouping execution module is used for converting the unique candidate identity into an identity confirmation signal and driving an external execution mechanism to complete grouping guide operation; and the data updating module is used for supplementing the obtained visual feature data to a pre-stored identity database. According to the method, the feature comparison range is pre-screened through the candidate identity numbers, and identity matching is performed by fusing multi-dimensional visual features such as orbit contours, nose print textures and ear blood vessel veins, so that the recognition accuracy and the environmental adaptability are improved.

Description

Technical Field

[0001] This invention relates to the field of grouping and guidance technology, and in particular to an intelligent identification and grouping guidance system suitable for camel milking parlors. Background Technology

[0002] A camel milking parlor is a fixed agricultural facility designed for large-scale, standardized camel milking operations. Its core functional areas include a waiting area, a milking operation area, a post-milking retention area, and multiple grouping areas physically separated by dedicated passageways and isolation fences. The facility is equipped with automated milking equipment, an individual camel identification system, behavioral guidance devices, and a central control unit. Through the coordinated operation of its structural layout and electromechanical systems, it achieves orderly guidance of the camel herd, accurate identification of individual camels, automatic grouping according to pre-set conditions (such as milk production stage and health status), and efficient and hygienic milking operations. The ultimate goal is to improve the efficiency, quality, and traceability management of camel milk production.

[0003] Existing camel milking parlor animal identification systems primarily rely on single biometric identification technologies, such as image recognition or radio frequency identification (RFID).

[0004] Such systems have significant drawbacks in practical applications: single identification technologies lack reliability in complex and ever-changing real-world scenarios. For example, relying solely on image recognition can lead to significant fluctuations in accuracy when camel postures are varied, lighting conditions are poor, or some facial features are obscured. Using RFID technology alone presents risks such as easily damaged electronic tags, blind spots in reading, and potential malicious disassembly. Therefore, there is an urgent need to provide an intelligent identification and grouping guidance system suitable for camel milking parlors to address these issues. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the lack of reliability of the existing single identification technology in complex and ever-changing real-world scenarios. For example, when relying solely on image recognition, the recognition accuracy is prone to significant fluctuations when camel postures are varied, lighting conditions are poor, or some facial features are obscured. On the other hand, using only RFID technology has the disadvantages of easily damaged electronic tags, blind spots in reading, and the risk of malicious disassembly. The present invention provides an intelligent identification and grouping guidance system suitable for camel milking parlors.

[0006] To solve the above-mentioned technical problems, one technical solution adopted by the present invention is to provide an intelligent identification and grouping guidance system suitable for camel milking parlors, including a data acquisition module, a feature processing module, a verification module, a grouping execution module and a data update module; The data acquisition module is used to acquire visual feature data of the camel through a preset sensor group and to acquire the camel's identification data through a radio frequency identification reader; The feature processing module is used to parse the identification data to generate candidate identity numbers, extract representative features from the visual feature data, and match the representative features with feature templates associated with the candidate identity numbers in a pre-stored identity database to generate an initial identity dataset. The verification module is used to monitor the initial identity dataset; when several candidate identities are detected in the initial identity dataset, a hierarchical verification process is initiated to filter and exclude the candidate identities in the initial identity dataset until only one candidate identity remains in the initial identity dataset; when only a single candidate identity is detected in the initial identity dataset, the single candidate identity is taken as the only candidate identity. The grouping execution module converts the unique candidate identity into an identity confirmation signal, retrieves the camel's physiological parameters based on the identity confirmation signal, generates grouping control instructions, and drives the external execution mechanism to complete the grouping guidance operation. The data update module supplements the pre-stored identity database with the visual feature data acquired during each identification process of the data acquisition module.

[0007] The present invention is further configured such that: the preset sensor group in the data acquisition module includes a high-resolution image sensor for acquiring multi-angle visual features of the camel's face and body surface, and a depth sensor for acquiring three-dimensional contour information of the camel's body shape; The visual feature data in the data acquisition module includes two-dimensional image features of the camel's face, such as the eye socket contour, nose texture, and ear blood vessels, acquired by a high-resolution image sensor, and three-dimensional point cloud data of the camel's back contour and torso, obtained by a depth sensor.

[0008] The present invention is further configured such that the step of generating the candidate identity number in the feature processing module is as follows: S1. The pixel area of ​​the orbital contour region is calculated as the first area, the coverage area of ​​the effective texture in the nasal texture region is calculated as the second area, and the cumulative area of ​​the significant blood vessel branches in the ear vascular network region is calculated as the third area. S2. Based on the preset feature weight coefficients, the first area, the second area and the third area are sequentially concatenated to form a preliminary original feature vector. The preliminary original feature vector is subjected to binary conversion and position rearrangement operations to generate an intermediate encoding sequence. S3. Combine the identification data to transform the intermediate encoding sequence, calculate and generate a digest value of a preset fixed length, and use the digest value as a candidate identity number.

[0009] The present invention is further configured such that the method for extracting the representative features in the feature processing module is as follows: S4. Based on the first area, the second area and the third area, calculate the first proportional coefficient between the first area and the second area, and simultaneously calculate the second proportional coefficient between the second area and the third area. Combine the first proportional coefficient and the second proportional coefficient into a primary geometric relationship feature vector, and use the three-dimensional point cloud data of the camel's back contour to calculate the radius of curvature of the torso. S5. Calculate the coefficient of variation of the primary geometric relationship feature vector and the torso curvature radius within a preset time window, filter and aggregate the primary geometric relationship feature vectors with the coefficient of variation lower than a first preset threshold or the torso curvature radius with the coefficient of variation lower than a second preset threshold, and generate an intermediate representative feature set, which is the representative feature. The feature template in the pre-stored identity database includes two-dimensional image features of camel eye socket contours, nose textures, and ear blood vessels based on historically collected data, as well as the primary geometric relationship feature vector and the torso curvature radius. The method for generating the initial identity dataset is as follows: S6. Compare the similarity between the intermediate representative feature set and the feature templates associated with the candidate identity number in the pre-stored identity database, calculate the matching degree score between each primary geometric relationship feature vector in the intermediate representative feature set and the associated feature vector in the feature template, generate a preliminary matching degree score set, and based on the preliminary matching degree score set and combined with a preset matching threshold, perform preliminary screening of the candidate identities associated with the candidate identity number to obtain a preliminary screened identity set; S7. For each candidate identity in the preliminary screening identity set, assign a preset confidence weight according to its matching score, and comprehensively consider the historical update frequency and the number of successful recognitions of the feature template to calculate the weighted confidence score of each candidate identity. Sort the identities in the preliminary screening identity set in descending order according to the weighted confidence score to generate an ordered candidate identity list, and take the candidate identity ranked first in the ordered candidate identity list as the primary candidate identity. S8. Monitor the ordered candidate identity list. When the weighted confidence score of the primary candidate identity exceeds a preset single identity confirmation threshold, the primary candidate identity is included as the only candidate identity in the initial identity dataset. When multiple candidate identities in the ordered candidate identity list have weighted confidence scores that all exceed a preset multi-identity inclusion threshold and their differences are within a preset allowable range, the multiple candidate identities are included in the initial identity dataset to complete the generation of the initial identity dataset.

[0010] The present invention is further configured such that the specific steps of the hierarchical verification process in the verification module include: when several candidate identities are detected in the initial identity dataset, the representative feature extracted for the first time is used as the starting verification point, and the first adjacent representative feature is scanned and determined as the first transitional verification feature point within its preset spatial neighborhood. Based on the matching result of the transitional verification feature point and the feature template, the candidate identities in the initial identity dataset are filtered and excluded. If the number of candidate identities in the initial identity dataset is still not unique after filtering and exclusion, the preset spatial neighborhood is expanded by a preset range, and the second adjacent representative feature is scanned and obtained as the second transitional verification feature point within the expanded preset spatial neighborhood for additional verification. This process is iterated until only one candidate identity remains in the initial identity dataset. The starting verification point, the first transitional verification feature point, and the second transitional verification feature point constitute the main verification path. At the same time, a backup verification path is set up. When the main verification path is blocked, the system switches to the backup verification path for re-verification.

[0011] The present invention is further configured such that the method for determining the initial extraction of the representative features is as follows: Q1. From the continuous multiple frames of visual feature data acquired in real time by the data acquisition module, extract the two-dimensional image features of the eye socket contour, nose texture and ear blood vessels of the camel's facial region in parallel, and simultaneously acquire the three-dimensional point cloud data of the camel's back contour and torso collected by the depth sensor at the same time stamp. Q2. Map the various two-dimensional image features and the three-dimensional point cloud data to a unified preset plane coordinate system according to their spatial relative positions and temporal synchronization. In the preset plane coordinate system, connect the eye socket contour, nose texture, ear blood vessel feature points and back contour key points generated by the three-dimensional point cloud data in sequence according to the feature proximity principle using preset logical connection lines to generate a static spatial feature topology map under a single frame image. Based on the static spatial feature topology map of multiple consecutive frames, dynamically link them according to the time sequence to form a dynamic verification trajectory. Q3. Calculate the variance of position coordinates and the rate of change of shape of each feature point in each dynamic verification trajectory within a preset time window. Based on the variance of position coordinates and the rate of change of shape, calculate a comprehensive stability score for each feature point included in each dynamic verification trajectory. Select the dynamic verification trajectory to which the feature point with the highest comprehensive stability score belongs as the benchmark verification trajectory. From the benchmark verification trajectory, the single feature point that is first clearly captured at the initial moment and has the highest comprehensive stability score is determined as the representative feature extracted for the first time.

[0012] The present invention is further configured such that the calculation step of the comprehensive stability score in step Q3 is as follows: Q301. For each feature point in each dynamic verification trajectory, extract its three-dimensional spatial coordinate sequence in consecutive frames within a preset time window, and calculate the displacement variance of each feature point in the X, Y, and Z axis directions as a primary displacement fluctuation index. At the same time, extract the curvature change value of each feature point at different times to form a curvature change sequence. Q302. Based on the primary displacement fluctuation index, calculate the positional stability coefficient of each feature point, and based on the curvature change sequence, calculate the morphological stability coefficient of each feature point. Weight the positional stability coefficient and the morphological stability coefficient to generate a single-dimensional comprehensive stability score for each feature point. Simultaneously, calculate the variance of the relative distance between different feature points within the same dynamic verification trajectory as the spatial relationship stability coefficient between the feature points. Then, perform a secondary fusion of the spatial relationship stability coefficient and the single-dimensional comprehensive stability score to form a multi-dimensional stability intermediate score for the feature points. Q303. Introduce a preset time decay weighting factor to the intermediate multidimensional stability score of all the feature points, integrate and sum the intermediate multidimensional stability scores of each feature point along the time dimension after the time decay weighting adjustment, and normalize the integration result with the total duration of the dynamic verification trajectory to calculate the comprehensive stability score of each feature point in the whole preset time window.

[0013] The present invention is further configured such that the steps for setting the backup verification path are as follows: Q4. Based on the benchmark verification trajectory, select other feature points from the dynamic verification trajectory besides the feature points already used in the main verification path to form a backup feature set. For each feature point in the backup feature set, calculate the path priority weight of each feature point according to its comprehensive stability score and its spatial distribution characteristics in the preset plane coordinate system. Arrange the feature points in the backup feature set in descending order according to the path priority weight to generate an ordered backup feature sequence. Q5. Continuously monitor the execution status of the main verification path. When the matching confidence of the first transition verification feature point or the second transition verification feature point is lower than the preset path switching threshold, immediately generate a path switching trigger signal. According to the path switching trigger signal, select the backup feature point with the highest priority weight from the ordered backup feature sequence as the backup starting verification point. Based on the spatial adjacency relationship between the backup starting verification point and the subsequent feature points in the ordered backup feature sequence, construct a backup verification path consisting of the backup starting verification point, the backup first transition verification feature point, and the backup second transition verification feature point connected in sequence.

[0014] The present invention is further configured such that: the specific content of the grouping execution module is as follows: the grouping execution module is used to input the identity confirmation signal into a pre-stored camel physiological parameter mapping table, query the milk production cycle stage and health status identifier uniquely corresponding to the candidate identity as key physiological parameters, and convert the key physiological parameters into the grouping control command containing the target channel number and action sequence according to the preset grouping strategy logic, and send the grouping control command to the external execution mechanism through the preset command driving interface to control the opening and closing of the channel gate and the changing of the guide indicator light, thereby completing the directional grouping guidance operation for the current camel.

[0015] The present invention is further configured such that: when the visual feature data is added to the pre-stored identity database in the data update module, after the clustering execution module successfully completes the clustering guidance operation, the specific camel identity record associated with the pre-stored identity database is locked according to the identity confirmation signal, and the intermediate representative feature set generated in this identification process and confirmed as valid by the verification module, along with the weighted confidence score, is stored as a new feature template in the specific camel identity record. At the same time, the existing historical feature templates in the pre-stored identity database are optimized and iterated according to the new feature template.

[0016] The beneficial effects of this invention are as follows: 1. This invention combines image sensor and radio frequency identification technology for multimodal data acquisition and feature processing, utilizes the candidate identity number provided by RFID to pre-screen the feature comparison range, and integrates multi-dimensional visual features such as eye socket contour, nose texture, and ear blood vessels for identity matching. This overcomes the limitations of single technology in complex scenarios such as camel posture changes, lighting changes, or partial occlusion, and significantly improves the accuracy and environmental adaptability of identification. 2. This invention evaluates matching results by introducing a weighted confidence score and designs a hierarchical verification process that includes a primary verification path and a backup verification path. When the confidence of the primary path verification is insufficient, it can intelligently switch to the backup path, which effectively reduces the risk of identification failure caused by electronic tag damage, reading blind spots or malicious disassembly, and ensures the reliability of the identity verification process and the continuous and stable operation of the system. 3. This invention updates the database with verified representative features and confidence scores as new templates after each successful grouping, and iteratively optimizes historical feature templates, so that the pre-stored identity database can continuously adapt to changes in the physiological characteristics of individual camels, thereby continuously improving the recognition performance and robustness of the system in long-term use. Attached Figure Description

[0017] Figure 1 This is a system flowchart of the present invention; Figure 2 This is a method flow for determining the first extraction of representative features of the present invention. Detailed Implementation

[0018] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0019] Please see Figures 1-2 A smart identification and grouping guidance system for camel milking parlors includes a data acquisition module, a feature processing module, a verification module, a grouping execution module, and a data update module. The data acquisition module is used to acquire visual feature data of the camel through a preset sensor group and to acquire the camel's identification data through a radio frequency identification reader; The feature processing module is used to parse the identification data to generate candidate identity numbers, extract representative features from the visual feature data, and match the representative features with the feature templates associated with the candidate identity numbers in the pre-stored identity database to generate the initial identity dataset. The verification module is used to monitor the initial identity dataset. When several candidate identities are detected in the initial identity dataset, a hierarchical verification process is initiated to filter and eliminate the candidate identities in the initial identity dataset until only one candidate identity remains in the initial identity dataset. When only a single candidate identity is detected in the initial identity dataset, the single candidate identity is taken as the only candidate identity. The grouping execution module converts the unique candidate identity into an identity confirmation signal, retrieves the camel's physiological parameters based on the identity confirmation signal, generates grouping control instructions, and drives the external actuator to complete the grouping guidance operation. The data update module supplements the pre-stored identity database with the visual feature data acquired during each data acquisition module's recognition process, thereby achieving continuous optimization of the feature database.

[0020] When a camel enters the identification area, the RFID reader can read the unique ID from its electronic tag (ear tag or leg band) in near real-time. This ID is directly linked to the camel's identity information. After obtaining this ID, the system does not directly use it as the final identity, but rather as a high-confidence prediction. Based on this prediction, it compares the feature templates of only this one or a few candidate camels in the database (considering the possibility of RFID misreads), instead of performing a full match with the features of all camels in the database. This improves the speed and efficiency of image recognition, especially suitable for high-throughput scenarios in milking parlors. If severe obstruction or damage prevents both the main and backup image recognition paths from determining a unique identity, the identity information provided by RFID can serve as a final, reliable fallback solution, ensuring that the system can still complete the identification and grouping tasks, avoiding system paralysis due to complete image recognition failure.

[0021] The preset sensor group in the data acquisition module includes a high-resolution image sensor for acquiring multi-angle visual features of the camel's face and body surface, a depth sensor for acquiring three-dimensional contour information of the camel's body shape; the preset sensor group also includes a photosensitive sensor for monitoring ambient light conditions, a temperature and humidity sensor for detecting ambient temperature and humidity, an infrared sensor for assisting identification under night or low light conditions, and a speed sensor for monitoring the camel's speed and direction. The visual feature data in the data acquisition module includes two-dimensional image features of the camel's face, such as the eye socket contour, nose texture, and ear blood vessels, acquired through a high-resolution image sensor, and three-dimensional point cloud data of the camel's back contour and torso, acquired through a depth sensor.

[0022] These characteristics should possess high uniqueness (able to distinguish different individuals) and stability (not changing drastically with posture, hair growth cycle, or minor soiling).

[0023] One embodiment of the present invention is as follows: the step of generating candidate identity numbers in the feature processing module is as follows: S1. The pixel area of ​​the eye socket contour region is calculated as the first area, the coverage area of ​​the effective texture in the nose texture region is calculated as the second area, and the cumulative area of ​​the significant blood vessel branches in the ear blood vessel network region is calculated as the third area. The calculation steps for the first, second, and third areas are as follows: First, the camel face image acquired by the high-resolution image sensor is preprocessed, and the eye socket contour region, the nasal texture region, and the ear blood vessel region are defined as three independent regions to be calculated. Next, within the eye socket contour region, the total number of all pixels belonging to the eye socket contour is counted, and the product of this total number and the actual physical area coefficient of a single pixel is used as the first area. Within the nasal texture region, sub-regions with significant texture features are identified using edge detection and texture analysis algorithms, and the coverage area of ​​these effective texture sub-regions is accumulated as the second area. Within the ear blood vessel region, the contours of all significant blood vessel branches are extracted using image segmentation technology, and the total area enclosed by these blood vessel branch contours is accumulated as the third area. S2. Based on the preset feature weight coefficients, the first area, the second area and the third area are concatenated in order to form a preliminary original feature vector. The preliminary original feature vector is subjected to binary conversion and position rearrangement operations to generate an intermediate encoding sequence. Bit order rearrangement operation: After generating the initial original feature vector, each numerical element in the vector is converted into a fixed-width binary string according to a preset binary conversion rule, and these binary strings are sequentially concatenated to form an initial binary sequence; then, according to a preset pseudo-random sequence related to the purpose of generating candidate ID numbers, the order of each bit in the initial binary sequence is rearranged. This process is called bit order rearrangement operation. The result is to generate an intermediate encoding sequence with higher disorder to enhance the distinguishability and security of subsequent processing. S3. Combine the identification data to transform the intermediate encoded sequence, calculate and generate a digest value of a preset fixed length, and use the digest value as a candidate identity number.

[0024] The calculation steps for a preset fixed-length digest value are as follows: After obtaining the intermediate encoded sequence, it is combined with the identification data obtained from the RFID reader as input and fed into a preset cryptographic hash function (such as a simplified variant of SHA-256) for iterative compression calculation. This calculation process maps input data of arbitrary length to a fixed-length and unique output data, which is the preset fixed-length digest value and is used as a candidate identity number for internal system use to ensure the uniqueness and consistency of the number. The method for extracting representative features in the feature processing module is as follows: S4. Based on the first area, the second area and the third area, calculate the first proportional coefficient between the first area and the second area, and at the same time calculate the second proportional coefficient between the second area and the third area. Combine the first proportional coefficient and the second proportional coefficient into a primary geometric relationship feature vector, and use the three-dimensional point cloud data of the camel's back contour to calculate the radius of curvature of the torso. The calculation steps for the trunk curvature radius are as follows: Based on the camel back contour and trunk 3D point cloud data obtained by the depth sensor, a representative ridge line is first extracted from the point cloud data along the camel spine, and a series of equally spaced sampling points are selected on the ridge line. For each sampling point, a local tangent plane is fitted using its neighboring point cloud data, and the radius of the approximate circle of the ridge line curvature at that point is calculated. This radius is the local trunk curvature radius at that point. Usually, the statistical average (such as the median) of the curvature radii at all sampling points is taken as the trunk curvature radius representing the degree of curvature of the entire trunk. S5. Calculate the coefficient of variation of the primary geometric relationship feature vector and the torso curvature radius within a preset time window, filter and aggregate primary geometric relationship feature vectors with a coefficient of variation lower than the first preset threshold or torso curvature radii lower than the second preset threshold, and generate an intermediate representative feature set, which is the representative feature. Preset time window: The start of the time window is usually triggered by the data acquisition module detecting the camel entering the effective recognition area. Its duration is preset based on the average walking speed of the camel when it normally passes through the recognition area and the required processing time. This is intended to ensure that a sufficient number of consecutive frames of visual feature data can be captured for subsequent coefficient of variation calculation and feature stability analysis. The steps for calculating the coefficient of variation are as follows: For the primary geometric relationship feature vector (which may contain multiple scalar components) and the trunk curvature radius, calculate the standard deviation of their values ​​at different time points within a preset time window; then, divide the calculated standard deviation by the average value of the corresponding feature within the preset time window, and the resulting ratio is the coefficient of variation for each feature; this coefficient is used to quantify the volatility of the feature in a short period of time. The lower the coefficient of variation, the more stable the primary geometric relationship feature vector or the trunk curvature radius is. The feature templates in the pre-stored identity database include two-dimensional image features and primary geometric relationship feature vectors based on historically collected camel eye socket contours, nasal textures, ear blood vessels, and torso curvature radius; The method for generating the initial identity dataset is as follows: S6. Compare the similarity between the intermediate representative feature set and the feature templates associated with the candidate identity number in the pre-stored identity database. Calculate the matching score between each primary geometric relationship feature vector in the intermediate representative feature set and the associated feature vector in the feature template. Generate a preliminary matching score set. Based on the preliminary matching score set and combined with the preset matching threshold, perform preliminary screening of the candidate identities associated with the candidate identity number to obtain a preliminary screening identity set. The steps for calculating the matching score are as follows: Each primary geometric relation feature vector in the intermediate representative feature set is compared with the corresponding historical feature vector stored in the feature template associated with the candidate identity number in the pre-stored identity database; the Euclidean distance or cosine similarity between the current feature vector and the template feature vector in each dimension is calculated, and these similarity measures are fused through a preset weighted summation formula to finally calculate a scalar value between 0 and 1. This scalar value is the matching score, which is used to represent the degree of similarity between the current feature and the template feature. The preset matching threshold is a critical similarity value determined in advance through training and optimization with a large amount of experimental data. When the calculated matching score is higher than the matching threshold, it is considered that the current feature is successfully matched with the feature template in the database, and the corresponding candidate identity is retained in the preliminary screening identity set. If it is lower than the matching threshold, the candidate identity is excluded from the current screening process. The matching threshold is a key parameter for controlling the recognition accuracy and false recognition rate. S7. For each candidate identity in the preliminary screening identity set, assign a preset confidence weight according to its matching score, and comprehensively consider the historical update frequency of the feature template and the number of successful recognitions to calculate the weighted confidence score of each candidate identity. Sort the identities in the preliminary screening identity set in descending order according to the weighted confidence scores to generate an ordered candidate identity list, and take the candidate identity ranked first in the ordered candidate identity list as the primary candidate identity. The preset confidence weights are a set of coefficients determined based on the matching score, which are divided into different intervals. The higher the matching score, the greater the corresponding confidence weight. The steps for calculating the weighted confidence score for each candidate identity are as follows: For each candidate identity in the initial screening set, firstly, multiply its matching score obtained during feature matching by a basic weighting factor calculated based on the number of successful historical recognitions of that candidate identity to obtain a basic confidence score; then, introduce an adjustment factor related to the historical update frequency of the feature template, giving positive adjustment to templates with a moderate update frequency and negative adjustment to templates that have not been updated for a long time or have been updated too frequently; finally, normalize the adjusted basic confidence score so that it falls within a preset score range, and the final value obtained is the weighted confidence score of that candidate identity. S8. Monitor the ordered candidate identity list. When the weighted confidence score of the primary candidate identity exceeds the preset single identity confirmation threshold, the primary candidate identity is included as the only candidate identity in the initial identity dataset. When multiple candidate identities in the ordered candidate identity list have weighted confidence scores that all exceed the preset multi-identity inclusion threshold and the differences between them are within the preset allowable range, the multiple candidate identities are included in the initial identity dataset to complete the generation of the initial identity dataset.

[0025] The preset single identity confirmation threshold is a high confidence threshold. When the weighted confidence score of the first-ranked primary candidate identity in the ordered candidate identity list exceeds this threshold, the system can directly confirm its uniqueness. The preset multi-identity inclusion threshold is a relatively low confidence threshold. When the weighted confidence scores of multiple candidate identities in the list all exceed this threshold, these identities will be included in the initial identity dataset for subsequent verification. The preset allowable range refers to the maximum allowable difference in weighted confidence scores between these candidate identities that exceed the multi-identity inclusion threshold. If the score difference is within this range, they are considered to be at a "difficult to distinguish" confidence level, and a tiered verification process needs to be initiated.

[0026] This embodiment significantly improves the accuracy and efficiency of camel identification in high-throughput scenarios in milking parlors by integrating RFID pre-identification with a precise matching and hierarchical verification mechanism based on multimodal biometrics (eye sockets, nasal prints, ear blood vessels, and torso shape). It enhances the system's fault tolerance and robustness when some features are occluded or changed by utilizing weighted confidence scoring and multi-threshold decision-making strategies. Furthermore, by continuously updating the visual feature data after successful identification to the feature database, adaptive optimization of feature templates is achieved, effectively addressing the physiological changes of individual camels over time, thereby ensuring the long-term stability of the system and the continuous improvement of the recognition rate.

[0027] One embodiment of the present invention is as follows: the specific steps of the hierarchical verification process in the verification module include: when several candidate identities are detected in the initial identity dataset, the representative feature extracted for the first time is used as the starting verification point, and the first adjacent representative feature is scanned and determined as the first transitional verification feature point in its preset spatial neighborhood. Based on the matching result of the transitional verification feature point and the feature template, the candidate identities in the initial identity dataset are screened and excluded. If the number of candidate identities in the initial identity dataset is still not unique after screening and exclusion, the preset spatial neighborhood is expanded by a preset range, and the second adjacent representative feature is scanned and obtained as the second transitional verification feature point in the expanded preset spatial neighborhood for additional verification. This process is repeated iteratively until only one candidate identity remains in the initial identity dataset. The starting verification point, the first transitional verification feature point, and the second transitional verification feature point constitute the main verification path. At the same time, a backup verification path is set up. When the main verification path is blocked, the system switches to the backup verification path for re-verification.

[0028] The method for determining the initial extraction of representative features is as follows: Q1. From the continuous multi-frame visual feature data acquired in real time by the data acquisition module, extract the two-dimensional image features of the eye socket contour, nose texture and ear blood vessels of the camel's facial region in parallel, and simultaneously acquire the three-dimensional point cloud data of the camel's back contour and torso collected by the depth sensor at the same time stamp. Standardize and preprocess the extracted features to eliminate scale and lighting differences, and temporarily store the preprocessed multi-dimensional feature data in the feature cache area. Q2. Map various two-dimensional image features and three-dimensional point cloud data to a unified preset plane coordinate system based on their spatial relative positions and temporal synchronization. Within the preset plane coordinate system, connect the eye socket contour, nose texture, ear blood vessel feature points and back contour key points generated from the three-dimensional point cloud data in sequence according to the feature proximity principle using preset logical connection lines to generate a static spatial feature topology map under a single frame image. Based on the static spatial feature topology map of multiple consecutive frames, dynamically link them according to the time series to form a dynamic verification trajectory. Preset logical connection lines: These are straight lines or curves that are virtually drawn to express the spatial relationship between feature points when generating a static spatial feature topology map. They are drawn based on the spatial proximity calculated by the Euclidean distance between feature points and follow a preset connection order from the feature points of the eye socket contour to the feature points of the nose texture, then to the feature points of the ear blood vessels, and finally to the key points of the back contour. The steps for forming a dynamic verification trajectory are as follows: under continuous timestamps, the static spatial feature topology map generated at each moment is arranged in chronological order; for each feature point, its position coordinates in the static spatial feature topology map at different moments are connected in time sequence to form the movement trajectory of the feature point over time; the individual movement trajectories of all feature points are superimposed and correlated in the spatiotemporal dimension to form a dynamic verification trajectory that reflects the dynamic changes of the overall characteristics of the camel. Q3. Calculate the variance of position coordinates and the rate of change of shape of each feature point in each dynamic verification trajectory within a preset time window. Based on the variance of position coordinates and the rate of change of shape, calculate the comprehensive stability score for each feature point contained in each dynamic verification trajectory. Select the dynamic verification trajectory to which the feature point with the highest comprehensive stability score belongs as the benchmark verification trajectory. From the benchmark verification trajectory, the single feature point that is first clearly captured at the initial moment and has the highest comprehensive stability score is determined as the representative feature extracted for the first time.

[0029] The calculation methods for position coordinate variance and morphological change rate are as follows: Position coordinate variance is calculated for a series of three-dimensional coordinate values ​​of a feature point within a preset time window. The variance of the values ​​on each of the X, Y, and Z axes is calculated, and then the arithmetic mean of the variances of the three axes is taken as the position coordinate variance of the feature point. Morphological change rate is characterized by calculating the absolute value of the change in contour curvature of the feature point between consecutive frames, and then taking the average of these absolute values ​​within the preset time window.

[0030] The calculation steps for the comprehensive stability score in step Q3 are as follows: Q301. For each feature point in each dynamic verification trajectory, extract its three-dimensional spatial coordinate sequence in consecutive frames within a preset time window, and calculate the displacement variance of each feature point in the X, Y, and Z axes as a primary displacement fluctuation index. At the same time, extract the curvature change value of each feature point at different times to form a curvature change sequence. Standardize the primary displacement fluctuation index and curvature change sequence to eliminate the influence of dimensions and provide standardized input data for subsequent stability index calculation. The steps for calculating the displacement variance of each feature point in the X, Y, and Z axes are as follows: First, obtain the three-dimensional coordinate sequence of a feature point in each frame within a preset time window; then, calculate the variance of the X coordinate value sequence, the variance of the Y coordinate value sequence, and the variance of the Z coordinate value sequence of the feature point; finally, use the calculated displacement variance in the X-axis direction, the displacement variance in the Y-axis direction, and the displacement variance in the Z-axis direction as the three components of the primary displacement fluctuation index. Q302. Based on the primary displacement fluctuation index, calculate the positional stability coefficient of each feature point. The positional stability coefficient is inversely proportional to the displacement variance. Based on the curvature change sequence, calculate the morphological stability coefficient of each feature point. The morphological stability coefficient is inversely proportional to the variance of the curvature change value. The positional stability coefficient and the morphological stability coefficient are weighted and fused to generate a single-dimensional comprehensive stability score for each feature point. At the same time, calculate the variance of the relative distance between different feature points within the same dynamic verification trajectory as the spatial relationship stability coefficient between feature points. The spatial relationship stability coefficient and the single-dimensional comprehensive stability score are then fused again to form a multi-dimensional stability intermediate score for the feature points. The calculation steps for the position stability coefficient of each feature point are as follows: The position stability coefficient is calculated by taking the reciprocal of the variance of the position coordinates and multiplying it by a preset scale normalization constant. This makes the feature point with the smaller the variance of the position coordinates (i.e., the more stable the position) have a larger position stability coefficient value. The calculation steps for the morphological stability coefficient of each feature point are as follows: The morphological stability coefficient is calculated by taking the reciprocal of the morphological change rate and multiplying it by another preset scale normalization constant, so that the feature point with the smaller morphological change rate (i.e., the more stable the morphology) has a larger morphological stability coefficient value. The specific steps for generating a single-dimensional comprehensive stability score for each feature point by weighting and fusing the positional stability coefficient and the morphological stability coefficient are as follows: assign a positional weight factor to the positional stability coefficient, assign a preset morphological weight factor to the morphological stability coefficient, and add the weighted positional stability coefficient and the weighted morphological stability coefficient together. The sum obtained is the single-dimensional comprehensive stability score. The specific steps for calculating the variance of the relative distance between different feature points within the same dynamic verification trajectory are as follows: At each sampling moment on the dynamic verification trajectory, the Euclidean distance between any two different feature points within the trajectory is calculated, thus obtaining a set of relative distance sequences that change over time; then, the variance of this relative distance sequence is calculated over the entire preset time window. This variance value is the spatial relationship stability coefficient between feature points. The smaller the coefficient value, the more stable the spatial layout relationship between feature points. The specific content of the intermediate score for multidimensional stability of feature points is to perform a secondary fusion of the spatial relationship stability coefficient and the single-dimensional comprehensive stability score. The inverse of the spatial relationship stability coefficient is multiplied by a preset relationship weight factor and then added to the single-dimensional comprehensive stability score. Alternatively, the single-dimensional comprehensive stability score is weighted and averaged with the spatial relationship stability coefficient after taking its inverse. The final value is the intermediate score for multidimensional stability. Q303. A preset time decay weighting factor is introduced into the multidimensional stability intermediate score of all feature points, so that the stability intermediate score calculated in recent frames has a higher weight, while the score weight of the earlier frames decreases frame by frame; the multidimensional stability intermediate score of each feature point after time decay weighting adjustment is integrated and summed along the time dimension, and the integration result is normalized with the total duration of the dynamic verification trajectory to calculate the comprehensive stability score of each feature point in the whole preset time window. This comprehensive stability score serves as the quantitative basis for screening the benchmark verification trajectory and determining the representative features extracted for the first time.

[0031] The preset time decay weight factor is a decreasing function coefficient that decreases as the timestamp is further away from the current time. Its purpose is to ensure that the feature stability contribution of recent frames is greater than that of earlier frames when calculating the comprehensive stability score. The calculation steps for the comprehensive stability score of each feature point within the entire preset time window are as follows: multiply the intermediate multidimensional stability score of the feature point in each frame within the preset time window by the preset time decay weight factor of the corresponding frame to obtain a series of time-weighted stability scores; then sum these weighted stability scores within the time window, and then normalize the sum by dividing the sum by the total number of frames (or total duration) of the preset time window. The final value is the comprehensive stability score of the feature point.

[0032] The steps for setting up the backup verification path are as follows: Q4. Based on the benchmark verification trajectory, select other (highly stable) feature points from the dynamic verification trajectory, excluding the feature points already used in the main verification path, to form a backup feature set. For each feature point in the backup feature set, calculate the path priority weight of each feature point according to its comprehensive stability score and its spatial distribution characteristics in the preset plane coordinate system. Arrange the feature points in the backup feature set in descending order according to the path priority weight to generate an ordered backup feature sequence. The steps for calculating the path priority weight of each feature point are as follows: use the comprehensive stability score of the feature point as the basic weight, and then multiply it by a distribution weight factor calculated based on the spatial distribution dispersion of the feature point and the selected path feature points in the preset planar coordinate system. The final product is the path priority weight of the feature point. The feature point with higher dispersion has a larger distribution weight factor. Q5. Continuously monitor the execution status of the main verification path. When the matching confidence of the first or second transitional verification feature point is lower than the preset path switching threshold, immediately generate a path switching trigger signal. Based on the path switching trigger signal, select the backup feature point with the highest priority weight from the ordered backup feature sequence as the backup starting verification point. Based on the spatial adjacency relationship between the backup starting verification point and subsequent feature points in the ordered backup feature sequence, construct a backup verification path consisting of the backup starting verification point, the backup first transitional verification feature point, and the backup second transitional verification feature point connected in sequence. After the backup verification path is successfully triggered and identity verification is completed for the first time, record the total time and final matching confidence of the verification process for this path as a path performance evaluation index. Compare and analyze the path performance evaluation index with the historical performance index of the main verification path. If the performance index of the backup path is better than that of the main path by a certain proportion, automatically adjust the path priority weight of the relevant feature points in the ordered backup feature sequence and update the optimized path configuration information to the system verification strategy library to realize the self-learning and continuous optimization closed loop of the backup verification path.

[0033] Path switching threshold: This refers to a critical confidence level value that is pre-calibrated and optimized using experimental data to measure the reliability of feature point matching on the main verification path. When the confidence level of matching the first or second transitional verification feature point calculated by the system in real time is lower than this threshold, it is determined that the reliability of the current verification link of the main verification path is insufficient, and the path switching mechanism is immediately triggered to stop the main path verification and start the process of re-verification by the backup verification path. The specific value of the path switching threshold is comprehensively set according to the system's tolerance for recognition error rate and the requirements for the continuity of the verification process.

[0034] In a specific example, the system presets a time window of 10 consecutive frames of data. When a camel enters the milking parlor's recognition area, the data acquisition module obtains its visual feature data through a high-resolution image sensor and a depth sensor. The feature processing module calculates the three-dimensional coordinates of the camel's nose mark center feature point within 10 frames, obtaining a variance of 0.8 for the X-axis displacement, 1.2 for the Y-axis, and 0.9 for the Z-axis. The arithmetic mean of the position coordinate variance is calculated to be 0.97, and the mean of its morphological change rate is calculated to be 0.15. The system takes a preset scale normalization constant of 10 and calculates the position stability coefficient as 10 / 0.97≈10.3 and the morphological stability coefficient as 10 / 0.15≈66.7. With a position weight factor of 0.4 and a morphological weight factor of 0.6, the weighted fusion yields a single-dimensional comprehensive stability score of 0.4*10.3+0.6*66.7≈44.7. Furthermore, the variance of the relative distance sequence between this feature point and another feature point (such as the orbital contour point) within 10 frames is calculated to be 0.5. The reciprocal of this variance is multiplied by the relationship weight factor of 0.5 to obtain the spatial relationship stability coefficient contribution value of 1.0. The multidimensional stability intermediate score is 45.7 after secondary fusion with the single-dimensional score. The system introduces a time decay weight factor (e.g., the weight of the most recent frame is 1.0, decreasing by 0.1 for each subsequent frame). The weighted score of 10 frames is integrated and summed to obtain 420. After normalization, the comprehensive stability score of the feature point is 42.0. This score is higher than the preset threshold of 35, so the feature point is selected as the starting point of the baseline verification trajectory. In the main verification path, when the matching confidence of the first transitional verification feature point (such as the ear blood vessel point) is calculated to be 0.75, which is lower than the path switching threshold of 0.8, the system immediately triggers a switch and selects the backup starting verification point with the highest priority weight from the ordered backup feature sequence (such as the back contour key point, with a comprehensive stability score of 38.0, a distribution weight factor of 1.2, and a path priority weight of 45.6), constructs the backup verification path, and successfully completes the verification. The total time for this backup path verification is 1.2 seconds, and the final matching confidence is 0.92, which is 20% better than the historical performance index of the main path (average time of 1.5 seconds and confidence of 0.85). The system automatically increases the path priority weight of the backup feature point in the sequence to achieve adaptive optimization of the verification path. Subsequently, the verification module confirms the camel's identity as the only candidate. The grouping execution module retrieves the pre-stored physiological parameter mapping table based on the identity confirmation signal, finds that the camel is in its peak lactation period, and generates a grouping control command. This command drives the external actuator to open the gate of the adjacent calf area and illuminate the guide indicator light, guiding the corresponding calf into the nursing area to stimulate the camel's lactation. After the stimulation is completed, the system controls the milking equipment to start the milking operation, realizing intelligent grouping guidance based on camel characteristics.

[0035] The pre-stored physiological parameter mapping table is a structured data table pre-created and stored in the system database. Its core function is to establish a mapping relationship between a camel's unique identity and its key physiological parameters. Once the system confirms a camel's identity through the identification and verification process, the grouping execution module immediately queries this table to obtain the physiological basis required for grouping decisions.

[0036] This embodiment achieves a precise quantitative assessment of the stability of camel biometrics by constructing a dynamic verification trajectory and calculating a multi-dimensional comprehensive stability score, thereby reliably determining the representative features extracted for the first time. By constructing a main verification path and a backup verification path through preset logical connection lines and intelligently switching when the confidence of the main path verification is insufficient, the robustness and continuity of the identity verification process are significantly improved in the case of partial feature occlusion or brief recognition failure, effectively ensuring the accuracy of camel identity recognition and the system operating efficiency in the high-throughput environment of the milking parlor.

[0037] One embodiment of the present invention is as follows: The specific content of the grouping execution module is as follows: The grouping execution module is used to input the identity confirmation signal into the pre-stored camel physiological parameter mapping table, query the milk production cycle stage and health status identifier that uniquely corresponds to the candidate identity as key physiological parameters, and convert the key physiological parameters into grouping control instructions containing target channel number and action sequence according to the preset grouping strategy logic. The grouping control instructions are sent to the external execution mechanism through the preset instruction drive interface to control the opening and closing of the channel gate and the changing of the guide indicator light, so as to complete the directional grouping guidance operation for the current camel.

[0038] The pre-stored camel physiological parameter mapping table refers to a database table that is pre-established and stored in the system. This table uses the camel's unique identifier as the primary key and records each camel's milk production cycle stage (such as early lactation, peak lactation, dry period), health status (such as healthy, under observation, requiring isolation), and other optional key physiological parameters (such as historical milk production records and immunization records). When the grouping execution module receives the identity confirmation signal, it quickly obtains the key physiological parameters that uniquely correspond to the currently identified camel by querying this mapping table, providing a decision-making basis for the subsequent generation of grouping control instructions.

[0039] This embodiment achieves automated and precise grouping guidance based on the physiological state of individual camels by querying a pre-stored physiological parameter mapping table and executing a preset grouping strategy, which significantly improves the efficiency of grouping operations and the level of management refinement in the milking parlor.

[0040] One embodiment of the present invention is as follows: When the visual feature data is supplemented to the pre-stored identity database in the data update module, after the grouping execution module successfully completes the grouping guidance operation, the specific camel identity record associated with the pre-stored identity database is locked according to the identity confirmation signal. The intermediate representative feature set and weighted confidence score generated in this identification process and confirmed as valid by the verification module are stored as new feature templates in the specific camel identity record. At the same time, the existing historical feature templates in the pre-stored identity database are optimized and iterated according to the new feature templates.

[0041] This embodiment automatically supplements the database with the valid feature data after successful identification, realizing continuous optimization and updating of feature templates, effectively adapting to the physiological changes of individual camels, and improving the recognition accuracy and stability of the system in long-term operation.

[0042] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An intelligent identification and grouping guidance system suitable for camel milking parlors, characterized in that: It includes a data acquisition module, a feature processing module, a verification module, a cluster execution module, and a data update module; The data acquisition module is used to acquire visual feature data of the camel through a preset sensor group and to acquire the camel's identification data through a radio frequency identification reader; The feature processing module is used to parse the identification data to generate candidate identity numbers, extract representative features from the visual feature data, and match the representative features with feature templates associated with the candidate identity numbers in a pre-stored identity database to generate an initial identity dataset. The verification module is used to monitor the initial identity dataset; when several candidate identities are detected in the initial identity dataset, a hierarchical verification process is initiated to filter and exclude the candidate identities in the initial identity dataset until only one candidate identity remains in the initial identity dataset; when only a single candidate identity is detected in the initial identity dataset, the single candidate identity is taken as the only candidate identity. The grouping execution module converts the unique candidate identity into an identity confirmation signal, retrieves the camel's physiological parameters based on the identity confirmation signal, generates grouping control instructions, and drives the external execution mechanism to complete the grouping guidance operation. The data update module supplements the pre-stored identity database with the visual feature data acquired during each identification process of the data acquisition module.

2. The intelligent identification and grouping guidance system for camel milking parlors according to claim 1, characterized in that: The preset sensor group in the data acquisition module includes a high-resolution image sensor for acquiring multi-angle visual features of the camel's face and body surface, and a depth sensor for acquiring three-dimensional contour information of the camel's body shape. The visual feature data in the data acquisition module includes two-dimensional image features of the camel's face, such as the eye socket contour, nose texture, and ear blood vessels, acquired by a high-resolution image sensor, and three-dimensional point cloud data of the camel's back contour and torso, obtained by a depth sensor.

3. The intelligent identification and grouping guidance system for camel milking parlors according to claim 2, characterized in that: The steps for generating the candidate identity number in the feature processing module are as follows: S1. The pixel area of ​​the orbital contour region is calculated as the first area, the coverage area of ​​the effective texture in the nasal texture region is calculated as the second area, and the cumulative area of ​​the significant blood vessel branches in the ear vascular network region is calculated as the third area. S2. Based on the preset feature weight coefficients, the first area, the second area and the third area are sequentially concatenated to form a preliminary original feature vector. The preliminary original feature vector is subjected to binary conversion and position rearrangement operations to generate an intermediate encoding sequence. S3. Combine the identification data to transform the intermediate encoding sequence, calculate and generate a digest value of a preset fixed length, and use the digest value as a candidate identity number.

4. The intelligent identification and grouping guidance system for camel milking parlors according to claim 3, characterized in that: The method for extracting the representative features in the feature processing module is as follows: S4. Based on the first area, the second area and the third area, calculate the first proportional coefficient between the first area and the second area, and simultaneously calculate the second proportional coefficient between the second area and the third area. Combine the first proportional coefficient and the second proportional coefficient into a primary geometric relationship feature vector, and use the three-dimensional point cloud data of the camel's back contour to calculate the radius of curvature of the torso. S5. Calculate the coefficient of variation of the primary geometric relationship feature vector and the torso curvature radius within a preset time window, filter and aggregate the primary geometric relationship feature vectors with the coefficient of variation lower than a first preset threshold or the torso curvature radius with the coefficient of variation lower than a second preset threshold, and generate an intermediate representative feature set, which is the representative feature. The feature template in the pre-stored identity database includes two-dimensional image features of camel eye socket contours, nose textures, and ear blood vessels based on historically collected data, as well as the primary geometric relationship feature vector and the torso curvature radius. The method for generating the initial identity dataset is as follows: S6. Compare the similarity between the intermediate representative feature set and the feature templates associated with the candidate identity number in the pre-stored identity database, calculate the matching degree score between each primary geometric relationship feature vector in the intermediate representative feature set and the associated feature vector in the feature template, generate a preliminary matching degree score set, and based on the preliminary matching degree score set and combined with a preset matching threshold, perform preliminary screening of the candidate identities associated with the candidate identity number to obtain a preliminary screened identity set; S7. For each candidate identity in the preliminary screening identity set, assign a preset confidence weight according to its matching score, and comprehensively consider the historical update frequency and the number of successful recognitions of the feature template to calculate the weighted confidence score of each candidate identity. Sort the identities in the preliminary screening identity set in descending order according to the weighted confidence score to generate an ordered candidate identity list, and take the candidate identity ranked first in the ordered candidate identity list as the primary candidate identity. S8. Monitor the ordered candidate identity list. When the weighted confidence score of the primary candidate identity exceeds the preset single identity confirmation threshold, the primary candidate identity is included in the initial identity dataset as the only candidate identity. When it is detected that the weighted confidence scores of multiple candidate identities in the ordered candidate identity list all exceed the preset multi-identity inclusion threshold and the difference between them is within the preset allowable range, the multiple candidate identities are jointly included in the initial identity dataset to complete the generation of the initial identity dataset.

5. The intelligent identification and grouping guidance system for camel milking parlors according to claim 4, characterized in that: The specific steps of the hierarchical verification process in the verification module include: when several candidate identities are detected in the initial identity dataset, the representative feature extracted for the first time is used as the starting verification point, and the first adjacent representative feature is scanned and determined as the first transitional verification feature point within its preset spatial neighborhood. Based on the matching result between the transitional verification feature point and the feature template, the candidate identities in the initial identity dataset are filtered and excluded. If the number of candidate identities in the initial identity dataset is still not unique after filtering and exclusion, the preset spatial neighborhood is expanded by a preset range, and the second adjacent representative feature is scanned and obtained as the second transitional verification feature point within the expanded preset spatial neighborhood for additional verification. This process is iterated until only one candidate identity remains in the initial identity dataset. The starting verification point, the first transitional verification feature point, and the second transitional verification feature point constitute the main verification path. At the same time, a backup verification path is set up. When the main verification path is blocked, the system switches to the backup verification path for re-verification.

6. The intelligent identification and grouping guidance system for camel milking parlors according to claim 5, characterized in that: The method for determining the initial extraction of the representative features is as follows: Q1. From the continuous multiple frames of visual feature data acquired in real time by the data acquisition module, extract the two-dimensional image features of the eye socket contour, nose texture and ear blood vessels of the camel's facial region in parallel, and simultaneously acquire the three-dimensional point cloud data of the camel's back contour and torso collected by the depth sensor at the same time stamp. Q2. Map the various two-dimensional image features and the three-dimensional point cloud data to a unified preset plane coordinate system according to their spatial relative positions and temporal synchronization. In the preset plane coordinate system, connect the eye socket contour, nose texture, ear blood vessel feature points and back contour key points generated by the three-dimensional point cloud data in sequence according to the feature proximity principle using preset logical connection lines to generate a static spatial feature topology map under a single frame image. Based on the static spatial feature topology map of multiple consecutive frames, dynamically link them according to the time sequence to form a dynamic verification trajectory. Q3. Calculate the variance of position coordinates and the rate of change of shape of each feature point in each dynamic verification trajectory within a preset time window. Based on the variance of position coordinates and the rate of change of shape, calculate a comprehensive stability score for each feature point included in each dynamic verification trajectory. Select the dynamic verification trajectory to which the feature point with the highest comprehensive stability score belongs as the benchmark verification trajectory. From the benchmark verification trajectory, the single feature point that is first clearly captured at the initial moment and has the highest comprehensive stability score is determined as the representative feature extracted for the first time.

7. The intelligent identification and grouping guidance system for camel milking parlors according to claim 6, characterized in that: The calculation steps for the comprehensive stability score in step Q3 are as follows: Q301. For each feature point in each dynamic verification trajectory, extract its three-dimensional spatial coordinate sequence in consecutive frames within a preset time window, and calculate the displacement variance of each feature point in the X, Y, and Z axis directions as a primary displacement fluctuation index. At the same time, extract the curvature change value of each feature point at different times to form a curvature change sequence. Q302. Based on the primary displacement fluctuation index, calculate the positional stability coefficient of each feature point, and based on the curvature change sequence, calculate the morphological stability coefficient of each feature point. Weight the positional stability coefficient and the morphological stability coefficient to generate a single-dimensional comprehensive stability score for each feature point. Simultaneously, calculate the variance of the relative distance between different feature points within the same dynamic verification trajectory as the spatial relationship stability coefficient between the feature points. Then, perform a secondary fusion of the spatial relationship stability coefficient and the single-dimensional comprehensive stability score to form a multi-dimensional stability intermediate score for the feature points. Q303. Introduce a preset time decay weighting factor to the intermediate multidimensional stability score of all the feature points, integrate and sum the intermediate multidimensional stability scores of each feature point along the time dimension after the time decay weighting adjustment, and normalize the integration result with the total duration of the dynamic verification trajectory to calculate the comprehensive stability score of each feature point in the whole preset time window.

8. The intelligent identification and grouping guidance system for camel milking parlors according to claim 7, characterized in that: The steps for setting up the backup verification path are as follows: Q4. Based on the benchmark verification trajectory, select other feature points from the dynamic verification trajectory besides the feature points already used in the main verification path to form a backup feature set. For each feature point in the backup feature set, calculate the path priority weight of each feature point according to its comprehensive stability score and its spatial distribution characteristics in the preset plane coordinate system. Arrange the feature points in the backup feature set in descending order according to the path priority weight to generate an ordered backup feature sequence. Q5. Continuously monitor the execution status of the main verification path. When the matching confidence of the first transition verification feature point or the second transition verification feature point is lower than the preset path switching threshold, immediately generate a path switching trigger signal. According to the path switching trigger signal, select the backup feature point with the highest priority weight from the ordered backup feature sequence as the backup starting verification point. Based on the spatial adjacency relationship between the backup starting verification point and the subsequent feature points in the ordered backup feature sequence, construct a backup verification path consisting of the backup starting verification point, the backup first transition verification feature point, and the backup second transition verification feature point connected in sequence.

9. The intelligent identification and grouping guidance system for camel milking parlors according to claim 8, characterized in that: The specific content of the grouping execution module is as follows: The grouping execution module is used to input the identity confirmation signal into a pre-stored camel physiological parameter mapping table, query the milk production cycle stage and health status identifier that uniquely corresponds to the candidate identity as key physiological parameters, and convert the key physiological parameters into the grouping control command containing the target channel number and action sequence according to the preset grouping strategy logic. The grouping control command is sent to the external execution mechanism through the preset command drive interface to control the opening and closing of the channel gate and the changing of the guide indicator light, so as to complete the directional grouping guidance operation for the current camel.

10. The intelligent identification and grouping guidance system for camel milking parlors according to claim 9, characterized in that: When the data update module supplements the visual feature data to the pre-stored identity database, after the grouping execution module successfully completes the grouping guidance operation, it locks the specific camel identity record associated with the pre-stored identity database based on the identity confirmation signal. The intermediate representative feature set generated during this identification process and confirmed as valid by the verification module, along with the weighted confidence score, are stored as new feature templates in the specific camel identity record. At the same time, the existing historical feature templates in the pre-stored identity database are optimized and iterated based on the new feature templates.

Images