A kind of multi-layer cage breeding scene-oriented breeder egg detecting robot and control method thereof

Abstract

The application discloses a kind of poultry inspection egg-collecting robot for multi-layer cage breeding scene and control method thereof, it is related to poultry breeding technical field, robot includes inspection egg-collecting robot body (1), mobile chassis (2), left front laser radar (3), right rear laser radar (4), left side extension bracket (5), right side extension bracket (6), left upper infrared camera (7), left lower infrared camera (8), right upper infrared camera (9), right lower infrared camera (10), industrial computer (11), chassis drive and CAN control unit (12), communication module (13), power supply and battery module (14), backfill base (15), cage position two-dimensional code label (16), ultrasonic obstacle avoidance sensor (17), inertial measurement unit (18), drive encoder (19) and touchable display screen (20). The application realizes the automation, precision and traceable acquisition of cage level egg production information.

Description

Technical Field

[0001] This invention relates to the field of poultry farming technology, specifically to a breeding poultry egg-detecting robot and its control method for multi-layer cage rearing scenarios. Background Technology

[0002] As poultry farming moves towards intensification and large-scale operations, multi-level cage rearing has become widely used in breeding poultry production due to its high space utilization and ease of centralized management. For breeding and production management, cage-level egg production data is crucial for evaluating individual egg production performance, conducting selective breeding, and implementing refined management. Therefore, continuous, accurate, and traceable recording of egg production in each cage is essential. Currently, egg production data acquisition in multi-level cage rearing still primarily relies on manual inspection, visual counting, and manual registration. This method is labor-intensive, involves poor working conditions, has high subjective error rates, and makes it difficult to establish a continuous and stable data loop.

[0003] To replace manual inspections, existing technologies have developed automatic counting solutions based on inspection chassis, monocular vision, or common image processing algorithms. However, under the real-world conditions of multi-story cage poultry houses, these solutions still have the following shortcomings:

[0004] Firstly, the poultry house is characterized by low light, dust, floating feathers, glare from cage bars, and partial obstruction. The hatching eggs are small in size and the QR codes have fine textures. Traditional single-task visual algorithms are prone to missed detections, false detections, and misinterpretations under low light, blurry conditions, and uneven lighting, making it difficult to simultaneously and stably complete hatching egg identification and cage location identification.

[0005] Secondly, the multi-layered cages are tall, have narrow passages, and limited viewing angles. A single camera or a small number of fixed-view cameras cannot simultaneously cover the upper, middle, and lower layers of cages and egg tanks, resulting in incomplete observations.

[0006] Third, when the robot is running in the poultry house passage with long, straight, and repetitive structures, the existing navigation methods that rely on single mapping and positioning or fixed waypoint configuration are prone to problems such as insufficient path reproduction accuracy, large cumulative positioning error, and difficulty in returning to the original inspection trajectory after obstacle avoidance.

[0007] Fourth, under the conditions of robot motion acquisition, the distance between hatching eggs and the QR codes of adjacent cages is close in the image. Especially when hatching eggs are located in the boundary area of ​​cages, existing technologies usually lack an egg cage attribution mechanism based on trajectory continuity, dynamic candidate constraints and perspective compensation, which can easily lead to cross-cage mismatch and duplicate counting.

[0008] Fifth, existing solutions often only provide identification results, lacking structured data recording, local caching, anomaly marking, breakpoint recovery, and retransmission mechanisms for inspection tasks, making it difficult to reliably store, trace, and verify data.

[0009] Therefore, there is an urgent need to propose a self-propelled poultry egg inspection robot and its control method that can adapt to the complex environment of low light in multi-layer cage rearing, take into account the overall hardware collaborative acquisition, robust dual-target detection under low light, precise matching of egg cages, high-precision autonomous navigation, and closed-loop data management, so as to realize the automated, accurate and traceable acquisition of cage-level egg production information. Summary of the Invention

[0010] To address the aforementioned shortcomings in existing technologies, this invention provides a poultry egg inspection robot and its control method for multi-layer cage rearing scenarios. This solution addresses the problems of small target size of hatching eggs, susceptibility of QR code targets to motion blur and uneven lighting, easy mismatch between eggs and cage positions, easy cumulative drift of robot pose in narrow and long passages, and difficulty in stable traceability of inspection results in multi-layer cage rearing scenarios.

[0011] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: a poultry egg-collecting robot for multi-layer cage rearing scenarios, comprising an inspection and egg-collecting robot body, a mobile chassis, a front left lidar, a rear right lidar, a left extension bracket, a right extension bracket, an upper left infrared camera, a lower left infrared camera, an upper right infrared camera, a lower right infrared camera, an industrial control computer, a chassis drive and CAN control unit, a communication module, a power supply and battery module, a recharge base, cage QR code labels, an ultrasonic obstacle avoidance sensor, an inertial measurement unit, a drive encoder, and a touch screen display.

[0012] The left front lidar is located on the left front side of the mobile chassis, and the right rear lidar is located on the right rear side of the mobile chassis, forming a complementary environmental perception structure. The left and right extension brackets are fixed to the left and right sides of the mobile chassis, respectively, and extend vertically. The upper left and lower left infrared cameras are mounted on the left extension bracket, and the upper right and lower right infrared cameras are mounted on the right extension bracket. The upper left and upper right infrared cameras are installed in a downward-looking manner, and the lower left and lower right infrared cameras are installed in an upward-looking manner, forming a composite field of view covering the front edge area of ​​the upper, middle, and lower cage positions, the egg tray area, and the cage position QR code label area. The cage position QR code label is located at the horizontal center position on the front side of each cage position.

[0013] The industrial control computer is connected to the front left lidar, the rear right lidar, the upper left infrared camera, the lower left infrared camera, the upper right infrared camera, the lower right infrared camera, the chassis drive and CAN control unit, the communication module, the ultrasonic obstacle avoidance sensor, the inertial measurement unit, and the drive encoder.

[0014] Furthermore, the recharge base is located at the end of the poultry house, the station area, or a preset docking area; the ultrasonic obstacle avoidance sensor is located on the front side or the two front corners of the mobile chassis; the inertial measurement unit and the drive encoder are installed inside the mobile chassis, wherein the inertial measurement unit is used to output the robot's heading angle, angular velocity, or attitude change information, and the drive encoder is used to output the displacement increment of the left and right drive wheels or the travel distance information.

[0015] The power supply and battery module is used to provide power to the mobile chassis, the left front lidar, the right rear lidar, the upper left infrared camera, the lower left infrared camera, the upper right infrared camera, the lower right infrared camera, the industrial control computer, and the communication module.

[0016] The present invention also adopts the following technical solution: a control method for a poultry egg-collecting robot for multi-layer cage rearing scenarios, comprising the following steps:

[0017] S1: The communication module receives the inspection task, the industrial control computer completes the initialization of sensor and chassis parameters, and loads path information, camera calibration parameters and counting judgment parameters;

[0018] S2: The chassis drive and CAN control unit control the mobile chassis to autonomously inspect along a preset path, collect environmental contour information through the left front lidar and right rear lidar, and simultaneously collect multi-layer cage position images through the left upper infrared camera, left lower infrared camera, right upper infrared camera and right lower infrared camera.

[0019] S3: The industrial control computer uses a built-in dual-target collaborative detection algorithm for hatching eggs and QR codes to simultaneously detect hatching egg targets and cage location QR code targets, thereby obtaining stable cage location identifiers and hatching egg trajectory information.

[0020] S4: The industrial control computer uses a built-in precise matching algorithm for egg cages based on time-series trajectory constraints and dynamic neighborhood affiliation. It uses stable cage position identifiers as cage position semantic anchors and egg trajectory information as matching objects to complete the unique affiliation determination between eggs and cage positions. When the stable matching conditions are met, it performs effective counting.

[0021] S5: The industrial control computer uses a built-in poultry house self-propelled inspection and navigation algorithm based on single-teaching path building, high-precision inertial reproduction and hierarchical obstacle avoidance and return to track. It continuously adjusts the robot's motion posture according to the current position, target path and obstacle status. After the obstacle is removed, the original inspection path is restored.

[0022] S6: After a single inspection channel or a single inspection task is completed, the industrial control computer organizes the inspection results into a structured record at the cage level and uploads it to the cloud server through the communication module. When all tasks are completed, the robot returns to the charging base to dock and recharge.

[0023] Furthermore, the dual-target collaborative detection algorithm for hatching eggs and QR codes includes the following steps:

[0024] S31: After distortion correction, temporal alignment and effective field of view cropping of the original multi-layer cage position image, shallow edge texture features and deep semantic features are extracted by a shared backbone network.

[0025] S32: Perform frequency domain feature enhancement processing and illumination sensing parallel fusion on shallow edge texture features and deep semantic features, and input the fused features into the dual-task detection head to output QR code candidate boxes and hatching egg candidate boxes respectively;

[0026] S33: After performing confidence screening and overlap suppression on the QR code candidate box and the hatching egg candidate box, the hatching egg detection box and the QR code detection box in the current frame are obtained;

[0027] S34: When the QR code detection frame simultaneously meets the trigger conditions of both global throttling and single trajectory throttling, the QR code image corresponding to the current trajectory is decoded, and the decoding result is written as a stable cage position identifier into the persistent decoding identifier register area, which remains valid throughout the life cycle of the trajectory.

[0028] Furthermore, the frequency domain feature enhancement includes:

[0029] The feature map is divided into local blocks, and a discrete cosine transform is performed on each local block to obtain the frequency domain feature map. ;

[0030]

[0031] in, It is the discrete cosine transform matrix. This is the feature map after local block reorganization. It is the transpose symbol;

[0032] Adaptive reweighting of different frequency bands in the frequency domain feature map is performed using convolutional layers, nonlinear activation layers, and the sigmoid function to obtain the frequency band weight matrix. The weighted frequency domain features are then inversely transformed back to the spatial domain to form the enhanced spatial features.

[0033]

[0034] in, The spatial features are recovered after frequency domain enhancement. For element-wise multiplication;

[0035] The weighted frequency domain features are inversely transformed back to the spatial domain and then residual-stacked with the original feature map.

[0036]

[0037] in, This is the feature map after residual superposition. For learnable balance coefficients, This is the original feature map.

[0038] Furthermore, the parallel fusion of illumination sensing includes:

[0039] The shallow edge texture features are concatenated with the upsampled deep semantic features and normalized, and then channel threshold denoising and spatial threshold denoising are performed in parallel. The channel threshold denoising is used to adaptively adjust the response intensity of each channel according to the image brightness level, and the spatial threshold denoising is used to suppress spatial noise caused by local reflections and uneven lighting.

[0040] In the channel threshold noise reduction branch, global average pooling and global max pooling are used together to generate the brightness perception threshold:

[0041]

[0042] in, The brightness perception threshold, It is the Sigmoid activation function. For learnable weights, It is a linear transformation function. This is a global average pooling operation. This is the normalized feature map. This is a global max pooling operation;

[0043] In the spatial threshold denoising branch, global and local brightness distributions are extracted along the spatial dimension, and a spatially adaptive threshold is generated:

[0044]

[0045] in, For spatial adaptive threshold, For convolution operations, For feature splicing operations, These are learnable weights.

[0046] Furthermore, the precise matching algorithm for egg cages based on temporal trajectory constraints and dynamic neighborhood affiliation includes the following steps:

[0047] S41: For each hatching egg trajectory, calculate the average Euclidean distance between it and each QR code trajectory in the candidate QR code set within the sliding time window;

[0048] S42: Introduce perspective compensation correction, determine the dynamic matching center by constructing an aspect ratio deviation curve, and define the effective matching range based on the center;

[0049] S43: Within the effective matching range, determine the optimal matching QR code according to the principle of minimum average Euclidean distance;

[0050] S44: Only when the optimal matching relationship is maintained within the sliding time window is the matching relationship confirmed as a valid egg cage matching result, and the hatching eggs are assigned to the corresponding cage position number.

[0051] Furthermore, the poultry house self-propelled inspection and navigation algorithm based on single-teaching path construction, high-precision inertial reproduction, and graded obstacle avoidance and return to the track includes the following steps:

[0052] S51: Manual remote control teaching records the inspection path, extracts key action points where the heading deviation exceeds the threshold, and compresses the continuous trajectory into segmented straight paths and turning sequences;

[0053] S52: Enhanced calibration inertial positioning is performed based on the drive encoder and inertial measurement unit, the heading angle is integrated to suppress cumulative drift, and a dual square figure-eight trajectory is used to correct the odometer system error;

[0054] S53: High-precision path tracking is achieved by synchronously converging lateral and heading deviations through feedback linearization;

[0055] S54: Divide obstacles into dynamic obstacles, small static obstacles, and large static obstacles, and implement real-time avoidance, circular detour, and parking waiting strategies respectively;

[0056] S55: Priority arbitration is performed on the path tracking control output and obstacle avoidance control output. Under normal conditions, path tracking control is in effect; when an obstacle is detected and the obstacle avoidance condition is triggered, obstacle avoidance control takes priority; after the obstacle avoidance is completed, it switches back to path tracking control.

[0057] S56: Calculate the consistency index between the reproduced path and the taught path, which is used to evaluate navigation quality, equipment debugging and acceptance.

[0058] Furthermore, when correcting the odometer system error, corrections are made for Type A errors caused by inconsistent wheel diameters and Type B errors caused by inconsistent wheel track, respectively:

[0059] The correction formula for Type A error is:

[0060]

[0061]

[0062] in, and The corrected wheel diameters for the left and right wheels. and These are the initially measured equivalent wheel diameters of the left and right wheels, respectively. To accumulate directional deviation, This represents the change from a theoretical perspective.

[0063] The correction formula for Type B error is:

[0064]

[0065] in, This is the corrected wheelbase. To initially measure the wheelbase, This represents the imbalance between clockwise and counterclockwise directions. and These represent the cumulative angle changes during clockwise and counterclockwise tests, respectively.

[0066] The beneficial effects of this invention are:

[0067] 1. Significantly improves the observation integrity in multi-cage scenarios and reduces missed detections and records.

[0068] This invention utilizes a mobile chassis, a front left lidar, a rear right lidar, a left-side extension bracket, a right-side extension bracket, an upper left infrared camera, a lower left infrared camera, an upper right infrared camera, and a lower right infrared camera to form a comprehensive data acquisition structure suitable for multi-layer cage rearing channels. The upper-layer cameras are installed from above, while the lower-layer cameras are installed from below, creating a composite field of view covering the frontal areas of the upper, middle, and lower cage positions, the egg tray area, and the QR code label area of ​​the cage positions. Compared to single-camera or a few fixed-view cameras, this invention reduces observation gaps caused by inter-layer obstruction, limited viewing angles, and local blind spots, thereby improving the completeness of multi-layer cage inspection data acquisition.

[0069] 2. Significantly improves the robustness of egg detection and cage location identification under low light, reflective, and shading conditions.

[0070] This invention constructs a detection network with a shared backbone and dual-task detection heads in an industrial control computer. In the feature extraction and fusion stage, it introduces a parallel fusion mechanism of frequency domain feature enhancement and illumination sensing, enabling the same network to simultaneously output candidate bounding boxes for hatching eggs and QR codes in a single forward inference. Simultaneously, it employs asynchronous decoding, global throttling, single-trajectory throttling, and a stable recognition strategy of "success on the first try, effective throughout." Therefore, compared to traditional single-task detection or single-frame decoding schemes, this invention reduces the probability of missed detections, false detections, and misinterpretations caused by weak light, local blurring, cage bar reflections, dust interference, and short-term occlusion, improving the stability of hatching egg and cage location identification, and reducing computational congestion caused by repeated decoding.

[0071] 3. Significantly improves the accuracy of the assignment between hatching eggs and cage locations, reducing mismatches across cages and duplicate counting.

[0072] This invention, after completing egg detection, QR code detection, QR code decoding, and multi-target tracking, constructs a dynamic candidate QR code set using the egg trajectory as anchor points. It then employs a combination of sliding time window average distance determination, dynamic matching center correction, and effective matching range constraints to complete cage matching. Only when the matching relationship remains stable within a continuous time window is a unique ownership confirmed and effective counting performed. Compared to single-frame nearest neighbor matching or fixed image center matching, this invention reduces the risk of ownership jumps under boundary region, perspective distortion, short-term jitter, and local occlusion conditions, thereby improving the uniqueness, accuracy, and traceability of cage-level egg production data.

[0073] 4. Significantly improves the stability, navigation accuracy, and path reproduction capability of self-propelled inspections in narrow, repetitive passageways of poultry houses.

[0074] This invention employs a navigation method that utilizes single-time manual teaching to define the path, enhanced calibration inertial positioning to estimate pose, feedback linearization control to track the path, and automatic return to the original path after handling graded obstacles. It does not rely on complex environment maps or manual point-by-point input of fixed waypoints. Compared to ordinary odometry navigation or navigation methods that solely rely on mapping, this invention improves path definition efficiency and reduces cumulative pose drift in scenarios with multi-layered cage housing with long passageways, significant repetitive structures, and limited lateral margins. Furthermore, it re-attaches to the original inspection path after obstacle avoidance. This allows the robot to continuously perform autonomous movement, track-following inspections, obstacle avoidance and track return, and continuous operations in a real poultry house environment, rather than merely replicating the trajectory in an ideal test path.

[0075] 5. Form a structured data closed loop from inspection to warehousing, reduce the burden of manual registration and improve the verifiability of results.

[0076] This invention generates cage-level structured records for results that meet the conditions of stable numbering, unique attribution, and valid counting during a single inspection. The records include at least the inspection time, cage number, egg production quantity, identification confidence level, corresponding trajectory identifier, and evidence frame index or key image information. When the communication link is normal, online uploading is performed via the communication module; when the communication link is abnormal, the industrial control computer first performs local caching, and then automatically re-uploads the data after the network is restored. Compared to manual inspection followed by manual registration or only outputting unstructured identification results, this invention reduces omissions, errors, and difficulties in subsequent traceability, improving the continuity, completeness, and management usability of inspection data.

[0077] 6. Significantly reduces on-site implementation complexity and overall usage costs.

[0078] This invention employs a mobile inspection system mounted on a mobile chassis, integrating a front left lidar, a rear right lidar, a multi-view infrared camera array, an industrial control computer, a communication module, and a recharge base into a single unit. This eliminates the need for numerous fixed acquisition devices and complex wiring at each cage or along an entire passageway. Furthermore, the shared backbone and dual-task detection heads reduce redundant calculations caused by deploying two separate detection models, while the asynchronous throttling decoding mechanism minimizes the impact of decoding on the real-time performance of the main thread. Therefore, this invention not only reduces the difficulty and complexity of poultry house renovations but also lowers computational resource consumption, maintenance costs, and reuse costs, thereby improving the system's engineering feasibility.

[0079] 7. It has good scene adaptability and scalability.

[0080] This invention allows for adjustments to the number, installation height, and angle of infrared cameras based on the number of poultry house floors, aisle width, and data collection requirements. It can also be extended to include environmental sensing units for correlation analysis between environmental information and cage-level egg production data. Furthermore, the data loop includes anomaly marking, anomaly list output, breakpoint resumption, and task archiving mechanisms. Therefore, this invention is not only applicable to multi-story cage rearing scenarios with different numbers of floors, layouts, and lighting conditions, but also facilitates subsequent expansion into environmental monitoring, anomaly early warning, and multi-task collaborative applications, demonstrating significant potential for widespread adoption.

[0081] In summary, this invention is not a simple superposition of existing detection, matching, navigation, and recording functions. Rather, it achieves automatic collection, accurate attribution, stable counting, and traceable management of breeding poultry egg production information in multi-layer cage rearing scenarios through the collaborative design of overall hardware layout, low-light dual-target detection, robust QR code decoding, dynamic egg cage matching, autonomous navigation control, and data closed-loop management. Therefore, it has strong advanced features, innovation, and practical application value. Attached Figure Description

[0082] Figure 1 This is a schematic diagram of a poultry egg-picking robot designed for multi-level cage rearing scenarios.

[0083] The components include: 1. The main body of the inspection and egg-collecting robot; 2. The mobile chassis; 3. The front left LiDAR; 4. The rear right LiDAR; 5. The left extension bracket; 6. The right extension bracket; 7. The upper left infrared camera; 8. The lower left infrared camera; 9. The upper right infrared camera; 10. The lower right infrared camera; 11. The industrial control computer; 12. The chassis drive and CAN control unit; 13. The communication module; 14. The power supply and battery module; 15. The recharge base; 16. The cage position QR code label; 17. The ultrasonic obstacle avoidance sensor; 18. The inertial measurement unit; 19. The drive encoder; and 20. The touch screen display.

[0084] Figure 2This is a flowchart of the control method for a poultry egg-picking robot designed for multi-layer cage rearing scenarios.

[0085] Figure 3 This is a flowchart of a dual-target collaborative detection algorithm for hatching eggs and QR codes.

[0086] Figure 4 The flowchart shows the precise matching algorithm for egg cages based on temporal trajectory constraints and dynamic neighborhood affiliation.

[0087] Figure 5 This is a flowchart of a poultry house self-propelled inspection and navigation algorithm based on single-teaching path construction, high-precision inertial reproduction, and graded obstacle avoidance and return to track. Detailed Implementation

[0088] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0089] Example 1, as Figure 1 As shown, a poultry egg-collecting robot for multi-layer cage rearing scenarios includes a robot body 1, a mobile chassis 2, a front left lidar 3, a rear right lidar 4, a left extension bracket 5, a right extension bracket 6, an upper left infrared camera 7, a lower left infrared camera 8, an upper right infrared camera 9, a lower right infrared camera 10, an industrial control computer 11, a chassis drive and CAN control unit 12, a communication module 13, a power supply and battery module 14, a recharging base 15, a cage QR code label 16, an ultrasonic obstacle avoidance sensor 17, an inertial measurement unit 18, a drive encoder 19, and a touch screen display 20.

[0090] The mobile chassis 2 serves as the platform for the entire machine, housing the industrial control computer 11, chassis drive and CAN control unit 12, communication module 13, power supply and battery module 14, and related installation structures. Under the control of the chassis drive and CAN control unit 12, it performs forward, backward, turning, U-turn, deceleration, obstacle stopping, and obstacle avoidance maneuvers. Preferably, the mobile chassis 2 is a differential wheel chassis to meet the requirements of low-speed, stable inspection and on-the-spot turning in narrow, long passageways of the poultry house.

[0091] The left front lidar 3 is located on the left front side of the mobile chassis 2, and the right rear lidar 4 is located on the right rear side of the mobile chassis 2. Preferably, they are located at approximately the same height plane, forming a complementary environmental perception structure for collecting information on the channel boundary, cage frame columns, obstacle outlines, and channel geometric features. The left extension bracket 5 and right extension bracket 6 are fixed to the left and right sides of the mobile chassis 2 respectively, extending vertically. The upper left infrared camera 7 and lower left infrared camera 8 are mounted on the left extension bracket 5, and the upper right infrared camera 9 and lower right infrared camera 10 are mounted on the right extension bracket 6. The upper left infrared camera 7 and upper right infrared camera 9 are installed in a downward-looking manner, while the lower left infrared camera 8 and lower right infrared camera 10 are installed in an upward-looking manner, forming a composite field of view covering the front edge area of ​​the upper, middle, and lower cage positions, the egg tray area, and the area of ​​the cage position QR code label 16. The cage position QR code label 16 is located at the horizontal center position on the front side of each cage position.

[0092] The industrial control computer 11 is connected to the front left lidar 3, the rear right lidar 4, the upper left infrared camera 7, the lower left infrared camera 8, the upper right infrared camera 9, and the lower right infrared camera 10, the chassis drive and CAN control unit 12, the communication module 13, the ultrasonic obstacle avoidance sensor 17, the inertial measurement unit 18, and the drive encoder 19. It is used to uniformly complete sensor data reception, pose estimation, navigation control, image reasoning and recognition, target tracking, robust QR code decoding, cage matching and counting, structured result generation, and task scheduling control. The communication module 13 is used to receive inspection tasks, upload inspection results, and interact with data on the cloud platform or user terminal.

[0093] The recharge base 15 is located at the end of the poultry house, the station area, or the preset docking area, and is used to automatically recharge the inspection and egg-collecting robot body 1 after the inspection is completed; the ultrasonic obstacle avoidance sensor 17 is located on the front side or the two front corners of the mobile chassis 2, and is used to sense dynamic obstacles or sudden intrusions in the robot's near field range, and to provide distance information for near-distance obstacle avoidance control during navigation; the inertial measurement unit 18 and the drive encoder 19 are installed inside the mobile chassis 2, wherein the inertial measurement unit 18 is used to output the robot's heading angle, angular velocity, or attitude change information, and the drive encoder 19 is used to output the displacement increment of the left and right drive wheels or the travel distance information. Together, they provide kinematic state input for path teaching, pose estimation, trajectory reproduction, and feedback control;

[0094] The power supply and battery module 14 is used to provide power to the mobile chassis 2, the left front lidar 3, the right rear lidar 4, the left upper infrared camera 7, the left lower infrared camera 8, the right upper infrared camera 9, the right lower infrared camera 10, the industrial control computer 11, and the communication module 13.

[0095] QR code labels 16 are fixedly installed on the front of each cage, preferably at the lateral center of the corresponding cage, to provide stable anchor points for cage identification, visual zoning, egg cage matching, and pose semantic correction. Thus, the front and rear dual lidars handle the tasks of channel boundary and obstacle contour perception, the multi-view infrared camera array handles cage image acquisition, the ultrasonic obstacle avoidance sensor 17 handles near-field obstacle perception, the inertial measurement unit 18 and drive encoder 19 handle motion state perception and pose estimation, the industrial control computer 11 handles navigation, identification, matching, and recording, the chassis drive and CAN control unit 12 handles motion execution, the communication module 13 handles task interaction and result uploading, and the power supply and battery module 14 and recharge base 15 jointly handle the overall power supply and battery life guarantee, thereby forming a complete inspection and egg-picking equipment suitable for multi-layer cage rearing scenarios.

[0096] In this embodiment, the robot can also be expanded to install an environmental sensing unit to synchronously collect environmental information such as poultry house temperature, humidity, carbon dioxide concentration and ammonia concentration; the number, installation height and installation angle of the infrared cameras can also be adjusted according to the number of poultry house floors, the width of the passage and the collection needs to adapt to the deployment requirements under different breeding scales and different cage layouts.

[0097] In this invention, the aforementioned hardware components are not set up independently, but rather work collaboratively around the operational chain of "navigation—collection—identification—matching—recording—recharging": After the communication module 13 receives the inspection task, the industrial control computer 11 initializes the parameters of each sensor and chassis, and the chassis drive and CAN control unit 12 control the mobile chassis 2 to travel along the inspection path; the left front lidar 3 and the right rear lidar 4 continuously collect environmental contour information for positioning, path tracking, and obstacle judgment; the upper left infrared camera 7, the lower left infrared camera 8, the upper right infrared camera 9, and the lower right infrared camera 10 continuously collect multi-view images for egg detection, QR code recognition, and subsequent egg cage matching; the industrial control computer 11 records the inspection results in a structured manner and uploads them to the cloud platform through the communication module 13; after the inspection is completed, the robot returns to the recharging base 15 for docking and charging, thus forming a fully automated inspection and egg-checking working mode for multi-layer cage rearing scenarios.

[0098] Example 2, as Figure 2 As shown, a control method for a poultry egg-collecting robot for multi-layer cage rearing scenarios includes the following steps:

[0099] S1: The communication module receives the inspection task, the industrial control computer completes the initialization of sensor and chassis parameters, and loads path information, camera calibration parameters and counting judgment parameters;

[0100] S2: The chassis drive and CAN control unit control the mobile chassis to autonomously inspect along a preset path, collect environmental contour information through the left front lidar and right rear lidar, and simultaneously collect multi-layer cage position images through the left upper infrared camera, left lower infrared camera, right upper infrared camera and right lower infrared camera.

[0101] S3: The industrial control computer uses a built-in dual-target collaborative detection algorithm for hatching eggs and QR codes to simultaneously detect hatching egg targets and cage location QR code targets, thereby obtaining stable cage location identifiers and hatching egg trajectory information.

[0102] To address the issues of missed detection, false detection, and unstable positioning caused by the small target size of hatching eggs, weak eggshell edges, fine QR code texture, blurred outlines under low light conditions, large day-night brightness variations, and significant interference from cage bar reflections and shadows in multi-layer cage rearing scenarios, a visual detection network with a shared backbone and dual-task detection heads is constructed within the industrial control computer. In the process of feature extraction and multi-scale feature fusion in the latter part of the backbone, frequency domain feature enhancement and illumination perception fusion mechanisms are introduced, enabling the same network to simultaneously output hatching egg candidate boxes and QR code candidate boxes in a single forward inference, providing a unified spatiotemporal coordinate benchmark for subsequent stable QR code recognition and egg cage matching.

[0103] like Figure 3 As shown, the dual-target collaborative detection algorithm for hatching eggs and QR codes includes the following steps:

[0104] S31: After distortion correction, temporal alignment and effective field of view cropping of the original multi-layer cage position image, shallow edge texture features and deep semantic features are extracted by a shared backbone network.

[0105] The upper left, lower left, upper right, and lower right infrared cameras acquire images according to a unified clock and transmit them to the industrial control computer. The industrial control computer first performs distortion correction, time alignment, and effective field-of-view cropping on the original images based on the camera calibration parameters, installation posture, and current inspection posture. Only the front edge of the cage, the egg trough area, and the area where the cage QR code label is located are retained. The background outside the corridor, invalid areas between layers, and strongly reflective areas on the edge of the cage are removed, reducing the interference of irrelevant background on subsequent inspections from the source.

[0106] The industrial control computer inputs the cropped image into a shared backbone network, which extracts shallow edge texture features and deep semantic features. A dual-task detection head is then set up after the backbone network, where the first detection branch outputs QR code candidate boxes and the second detection branch outputs hatching egg candidate boxes. The two branches share the same set of basic features and do not perform two separate sets of forward inference, thus maintaining a unified coordinate system and timestamp for both hatching egg and QR code detection results while ensuring real-time performance.

[0107] S32: Perform frequency domain feature enhancement processing and illumination sensing parallel fusion on shallow edge texture features and deep semantic features, and input the fused features into the dual-task detection head to output QR code candidate boxes and hatching egg candidate boxes respectively;

[0108] To address the issues of blurred egg outlines, unclear QR code positioning corners, and broken textures within QR code modules under low-light conditions, the industrial control computer performs frequency domain enhancement processing on the multi-scale feature maps output from the back-end of the backbone network. Specifically, the feature map is divided into several non-overlapping 8×8 local blocks, and a discrete cosine transform is performed on each local block to obtain the frequency domain feature map. ;

[0109]

[0110] in, It is the discrete cosine transform matrix. This is the feature map after local block reorganization. It is the transpose symbol;

[0111] Adaptive reweighting of different frequency bands in the frequency domain feature map is performed using convolutional layers, nonlinear activation layers, and the sigmoid function to obtain the frequency band weight matrix. The weighted frequency domain features are then inversely transformed back to the spatial domain to form the enhanced spatial features.

[0112]

[0113] in, The spatial features are recovered after frequency domain enhancement. This is an element-wise product (Hadamard product).

[0114] The weighted frequency domain features are inversely transformed back to the spatial domain and then residual-stacked with the original feature map.

[0115]

[0116] in, This is the feature map after residual superposition. For learnable balance coefficients, This is the original feature map.

[0117] This processing enhances the edges of hatching eggs, the corner points of QR codes, and the texture of internal modules of QR codes under low light and slight blur conditions, thereby improving the stability of small target detection.

[0118] To address the significant brightness variations caused by the alternation of natural light and infrared supplemental lighting within poultry houses, as well as the unstable foreground-background contrast due to cage bar reflections, localized shadows, and overexposure noise, the industrial control computer performs illumination sensing fusion on features at different scales. Specifically, it first fuses deep features... Upsampling to shallow features At the same resolution, the data is stitched together along the channel direction and normalized. Then, channel threshold denoising and spatial threshold denoising are performed in parallel, instead of using two consecutive thresholding processes. In practice, it has been proven that the parallel structure can retain more key structural information while suppressing redundant information.

[0119] In the channel threshold noise reduction branch, global average pooling and global max pooling are used together to generate the brightness perception threshold:

[0120]

[0121] in, The brightness perception threshold, It is the Sigmoid activation function. For learnable weights, It is a linear transformation function. This is a global average pooling operation. This is the normalized feature map. This is a global max pooling operation; this branch is used to adaptively adjust the response intensity of each channel according to the current image brightness level, so that the contrast relationship between the bright background and the weak texture target is redistributed.

[0122] In the spatial threshold denoising branch, global and local brightness distributions are extracted along the spatial dimension, and a spatially adaptive threshold is generated:

[0123]

[0124] in, For spatial adaptive threshold, For convolution operations, For feature splicing operations, These are learnable weights. This branch is used to suppress spatial noise caused by local high reflectivity, cage shadows, and uneven lighting. After parallel channel thresholding and spatial thresholding, the results are fused with the depthwise separable convolution output to obtain multi-scale detection features that combine edge details, semantic information, and brightness adaptability.

[0125] After obtaining the candidate egg frame, the industrial control computer further extracts the geometric size features, shell grayscale distribution features, surface texture features, edge continuity features, reflectivity features, and local contrast features with the background of the candidate target. Combined with the current image brightness status, local occlusion status, and motion blur degree, the computer makes a joint determination of the category and recognition difficulty of the current egg target.

[0126] For different types of eggs, the system does not adopt a uniform fixed threshold and a uniform fixed judgment strategy. Instead, it adaptively adjusts the confidence threshold, non-maximum suppression threshold, category decision weight, review time window length, and subsequent trajectory stability confirmation conditions according to the joint judgment results, so as to improve the recognition accuracy and output stability of different types of eggs under complex poultry house conditions.

[0127] When the category determination result of a single frame fluctuates, the system further combines the category results in the continuous frames of the trajectory to perform time-series voting or stability confirmation. Only when the category results remain consistent within the preset time window or the dominant advantage meets the conditions will the final category result of the trajectory be output.

[0128] The industrial control computer feeds the multi-scale features, after frequency domain enhancement and illumination sensing fusion, into the dual-task detection head, simultaneously outputting egg detection results and QR code detection results at at least three scales. A higher-resolution feature layer is used to improve the recall rate of small-sized QR codes and distant eggs; a lower-resolution feature layer is used to maintain the detection stability of near-range targets and partially occluded targets. This forms a complete detection chain: "shared backbone feature extraction—frequency domain enhancement for detail filling—illumination sensing for noise suppression—dual-task multi-scale output."

[0129] During the training phase, to improve the network's learning ability for samples with low light, reflection, occlusion, perspective compression, boundary cutting, and motion blur, the industrial control computer optimized the dual-task detection network using a joint loss function:

[0130]

[0131] in, For bounding box regression loss, Confidence loss for the target The category loss is used. The bounding box regression term uses CIoU to constrain the position, scale, and aspect ratio relationship between candidate boxes and ground truth boxes; the classification and confidence terms use focal loss to ensure that the training process focuses on difficult samples such as low-light QR codes, partially occluded QR codes, small-sized eggs, and edge-mounted eggs. The focal loss is written as:

[0132]

[0133] in, Predict the probability for the true class. This is the category balance coefficient. This is the focus coefficient for difficult samples. Through this training method, the network is no longer dominated by a large number of easily classified background samples, but instead allocates more learning capacity to difficult targets in actual inspection scenarios. Furthermore, for situations where the number of egg samples of different categories is uneven or the recognition difficulty varies greatly, differentiated training weights are set for each category of samples, so that the network can allocate higher learning attention to easily confused categories, low-proportion categories, and complex imaging categories during the training process.

[0134] S33: After performing confidence screening and overlap suppression on the QR code candidate box and the hatching egg candidate box, the hatching egg detection box and the QR code detection box in the current frame are obtained;

[0135] After the industrial control computer performs confidence filtering and overlap suppression on the candidate bounding boxes output by the dual-task detection heads, it obtains the set of egg detection boxes and the set of QR code detection boxes in the current frame. The QR code detection boxes are directly sent to the subsequent asynchronous decoding process, while the egg detection boxes are directly sent to the subsequent target tracking and egg cage matching process. Since both the eggs and QR codes originate from the same detection network and the same unified feature representation at the same time, the subsequent trajectory association, QR code attribution, and egg cage matching are all based on the same coordinate reference, reducing position drift and temporal misalignment caused by asynchronous inference from multiple models.

[0136] S34: When the QR code detection frame simultaneously meets the trigger conditions of both global throttling and single trajectory throttling, the QR code image corresponding to the current trajectory is decoded, and the decoding result is written as a stable cage position identifier into the persistent decoding identifier register area, which remains valid throughout the life cycle of the trajectory.

[0137] To prevent computational congestion caused by repeated QR code decoding, the system employs two levels of triggering conditions: global throttling and single-track throttling. Throttling only applies when a candidate QR code simultaneously meets the following conditions:

[0138] ,

[0139] Only then is it allowed to [do something]. The QR code image corresponding to each trajectory is submitted to a background thread for decoding; among them... The minimum time interval for global decoding. For the minimum attempt interval of a single trajectory, Let i be the timestamp of the first decoding attempt triggered on the i-th trajectory. This is the timestamp corresponding to the current moment. Through the aforementioned throttling mechanism, system lag can be avoided while ensuring a high success rate in recognition.

[0140] Once a valid decoding result is obtained for the same trajectory, the corresponding decode_id is written to the persistent decoding identifier register and remains valid throughout the trajectory's lifecycle, without being cleared due to partial occlusion, blurring, or angle changes in subsequent frames. Even if decoding fails temporarily in subsequent frames, the system still uses the written decode_id as a stable cage position identifier for the trajectory, thus forming a "success once, valid throughout" QR code recognition mechanism.

[0141] In the initial stage of trajectory establishment, to shorten the waiting time and improve the recognition probability of the first frame, the system allows the trajectory thread to trigger an early decoding of the newly created trajectory. After successful early decoding, the decode_id is immediately written back to the corresponding trajectory and kept consistent with the asynchronous decoding result in the background. If the background thread subsequently obtains the same result, the original identifier remains unchanged. If the current frame cannot be successfully decoded but the relationship between the adjacent identified cage position numbers and the current pose is clear, the cage position identifier of the current trajectory is auxiliaryly verified by combining the neighboring number inference result, thereby improving the numbering integrity during continuous inspection.

[0142] For the detected candidate hatching egg frames, the system outputs them to the subsequent target tracking and cage matching process. For QR code trajectories with stable decode_ids, the system uses them as semantic anchors for the current cage position and inputs them into the subsequent virtual detection area generation process. Only when both the hatching egg trajectory and the stable QR code identifier are valid are subsequent cage matching and counting processes allowed, thus preventing undecoded QR codes, momentarily miscoded QR codes, or cross-cage jitter from affecting the hatching egg attribution results.

[0143] After completing the above-mentioned target detection of hatching eggs, QR code detection, QR code decoding and multi-target tracking, the industrial control computer further performs the matching between hatching eggs and breeding cages to determine the unique cage number corresponding to each hatching egg.

[0144] S4: The industrial control computer uses a built-in precise matching algorithm for egg cages based on time-series trajectory constraints and dynamic neighborhood affiliation. It uses stable cage position identifiers as cage position semantic anchors and egg trajectory information as matching objects to complete the unique affiliation determination between eggs and cage positions. When the stable matching conditions are met, it performs effective counting.

[0145] To address the mismatching of egg cages caused by camera tilt, local perspective distortion, approximate distances in boundary areas, short-term occlusion, and motion blur in multi-layer cage rearing scenarios, the industrial control computer establishes a set of hatching egg trajectories and a set of QR code trajectories in the same time-series link. It also constructs a dynamic candidate QR code set using each hatching egg trajectory as an anchor point and completes the egg cage matching using a method of "dynamic neighborhood attribution matching + average distance determination + perspective compensation correction".

[0146] like Figure 4As shown, the precise matching algorithm for egg cages based on temporal trajectory constraints and dynamic neighborhood affiliation includes the following steps:

[0147] S41: For each hatching egg trajectory, calculate the average Euclidean distance between it and each QR code trajectory in the candidate QR code set within the sliding time window;

[0148] The industrial control computer receives the egg detection results, QR code detection results, QR code decoding results, and corresponding multi-target tracking results from each camera at the current moment, and generates sets of egg trajectories for each. and QR code trajectory set Among them, the first The trajectory of the seed egg at any moment The center coordinates are , No. The QR code trajectory at any time The center coordinates are Only QR code trajectories that have obtained valid QR code numbers, and QR code trajectories that, although not yet successfully decoded directly, have had their temporary numbers inferred from adjacent recognized QR codes, are included in the set of matchable cage positions.

[0149] When the system detects a newly appearing hatching egg trajectory, it does not directly perform a one-time nearest neighbor match with all QR codes in the entire map. Instead, starting from the moment the hatching egg trajectory is established, it builds a candidate QR code set for that trajectory. This candidate QR code set is updated in real-time in subsequent frames according to the trajectory's status, thus forming a dynamic neighborhood affiliation set for each individual hatching egg trajectory. Each hatching egg trajectory independently maintains its candidate QR code set to ensure uniqueness and real-time matching when multiple hatching eggs appear in parallel.

[0150] For the For each seed egg trajectory, calculate its Euclidean distance to each QR code trajectory in the candidate QR code set within the current frame:

[0151]

[0152] in, Indicates time Next The trajectory of the seed egg and the first The spatial distance between the QR code tracks.

[0153] To avoid instantaneous changes in matching results caused by single-frame detection jitter, slight occlusion, or local blurring, the system does not directly output the matching result using the minimum distance in a single frame. Instead, it outputs the result over a length of [missing information]. The average Euclidean distance is obtained by averaging the distances within a sliding time window:

[0154]

[0155] in, Indicates the number of times within the current sliding time window. The trajectory of the seed egg and the first The average distance between the QR code trajectories. The system determines the optimal matching QR code for the current egg trajectory according to the principle of minimum average Euclidean distance:

[0156]

[0157] in, For a moment The set of valid candidate QR codes, For the first The optimal QR code trajectory corresponding to the trajectory of the hatching egg.

[0158] S42: Introduce perspective compensation correction, determine the dynamic matching center by constructing an aspect ratio deviation curve, and define the effective matching range based on the center;

[0159] When the hatching egg is located in the central area of ​​the cage, the minimum average Euclidean distance directly reflects the true spatial proximity between the egg and its associated QR code. However, when the egg is located at the edge of the cage, the average distances from adjacent QR codes to the egg may be similar, leading to inconsistent attribution if the minimum distance is the sole determining factor. Therefore, the system further performs perspective distortion assessment on each QR code detection frame within the current frame, calculating the aspect ratio deviation of the detection frames.

[0160]

[0161] in, and They represent the first The width and height of the QR code detection frame. The QR code itself is approximately square. The closer the aspect ratio of the detection frame is to 1, the less the QR code's location is affected by perspective distortion.

[0162] The industrial control computer constructs an aspect ratio deviation curve based on the relationship between the horizontal coordinates of the QR code detection box within the current frame and the corresponding aspect ratio deviation. It then performs a fifth-order polynomial fitting on this curve to obtain:

[0163]

[0164] in, To fit the polynomial, As variables, , , , , , The correlation coefficient.

[0165] Take the horizontal coordinate corresponding to the minimum point of the fitted curve. As the dynamic matching center in the current field of view, that is:

[0166]

[0167] The dynamic matching center is located on the opposite side of the camera tilt direction and is used to compensate for lateral projection compression errors caused by the camera tilting to the left or right.

[0168] S43: Within the effective matching range, determine the optimal matching QR code according to the principle of minimum average Euclidean distance;

[0169] With dynamic matching center Based on this, establish effective matching ranges on both its left and right sides. Set the matching radius to [value]. Then the QR code trajectory participating in the egg cage matching in the current frame satisfies:

[0170]

[0171] in, This refers to the set of QR code trajectories within the valid matching range of the current frame. In implementation, Calibration can be performed based on the camera mounting height, lens field of view, and cage width; in a preferred embodiment, Take 300 pixels. Through the above processing, the system no longer uses a fixed image center as a matching reference, but automatically corrects the matching center according to the current viewpoint, thereby reducing the probability of mismatches in boundary areas and at tilted viewpoints.

[0172] S44: Only when the optimal matching relationship is maintained within the sliding time window is the matching relationship confirmed as a valid egg cage matching result and the hatching eggs assigned to the corresponding cage position number.

[0173] Obtain the optimal matching QR code Then, the system reads the corresponding QR code number as the cage number of the current hatching egg trajectory. If multiple hatching eggs exist in the same cage, each egg trajectory independently completes the candidate set update, distance calculation, and attribution determination. If the same hatching egg experiences partial occlusion, outline changes, or slight position fluctuations in consecutive frames, the system continues to use the dynamic candidate set and average distance results of that trajectory for updates without re-initializing the attribution relationship.

[0174] To ensure stable output results, the system only recognizes a matching relationship as a valid cage match if a QR code trajectory consistently maintains the minimum average Euclidean distance advantage within the current sliding time window. Once confirmed, the "Egg Trajectory Number - QR Code Number - Cage Number" is written into the matching result table for this inspection cycle, and is not counted repeatedly within the lifecycle of that egg trajectory. This ensures that each egg belongs to only one cage, and the same egg is recorded only once within the same inspection cycle.

[0175] The industrial control computer outputs the confirmed egg cage matching results to the subsequent counting and data storage process. The output content includes at least the hatching egg trajectory number, corresponding QR code number, cage number, matching establishment time, matching establishment image frame, hatching egg center coordinates, and QR code center coordinates. This forms a complete closed loop of "hatching egg detection - QR code recognition - time-series tracking - dynamic matching - unique attribution - counting and storage".

[0176] S5: The industrial control computer uses a built-in poultry house self-propelled inspection and navigation algorithm based on single-teaching path building, high-precision inertial reproduction and hierarchical obstacle avoidance and return to track. It continuously adjusts the robot's motion posture according to the current position, target path and obstacle status. After the obstacle is removed, the original inspection path is restored.

[0177] According to the inspection and egg-collecting robot body, mobile chassis, left front lidar, right rear lidar, ultrasonic obstacle avoidance sensor, inertial measurement unit, drive encoder and industrial control computer described in Example 1, the robot's navigation control in the multi-layer cage-raising scenario does not adopt the synchronous positioning and mapping method that relies on the environmental map, nor does it adopt the path configuration method of manually measuring coordinates point by point and pre-writing fixed waypoints. Instead, it adopts a navigation control method of "single manual teaching to quickly define the path - enhanced calibration inertial positioning to estimate pose in real time - feedback linearization to track high-precision trajectory - graded obstacle avoidance and return to the original path" to solve the problems of strong repetitive structure, narrow passage, light fluctuation, dust interference, low path definition efficiency and difficulty in re-attaching to the original inspection trajectory after obstacle avoidance in multi-layer cage-raising channels. This navigation control method is not a navigation method for general open environments, but a self-propelled inspection navigation method for multi-layer cage poultry houses with long, straight, narrow, repetitive structures, strong dust interference, and small lateral margins. Its core is to enable the robot to continuously, stably, and autonomously complete inspection, obstacle avoidance, path return, and operation reproduction within the poultry house passage.

[0178] like Figure 5 As shown, the poultry house self-propelled inspection and navigation algorithm based on single-teaching path construction, high-precision inertial reproduction, and graded obstacle avoidance and return to the track includes the following steps:

[0179] S51: Manual remote control teaching records the inspection path, extracts key action points where the heading deviation exceeds the threshold, and compresses the continuous trajectory into segmented straight paths and turning sequences;

[0180] Before navigation begins, the operator places the robot at the starting position of the inspection path and initializes its current pose using the inertial measurement unit and drive encoder. Subsequently, the operator remotely drives the robot along the desired inspection route. During the teaching process, the industrial control computer simultaneously records the robot's starting point, trajectory, turning actions, and ending point. The recorded teaching trajectory is not directly saved as a high-density discrete point sequence, but rather compressed in real-time into several key action points and adjacent line segments, thereby reducing storage burden and improving the efficiency of subsequent path reproduction.

[0181] During the teaching process, the industrial control computer reads the heading angle output by the inertial measurement unit in real time. When it detects that the deviation between the current heading and the initial heading of the current line segment exceeds a set threshold, the position is determined as a critical motion point. The determination criteria for critical motion points are written as follows:

[0182]

[0183] in, The heading angle at the current moment. This is the starting heading angle for the current straight segment. The threshold for steering recognition is set. Each key action point is recorded as:

[0184]

[0185] in, Here are the coordinates of the point. This represents the cumulative turning angle corresponding to that point. In this way, the continuous teaching trajectory is converted into a segmented straight path and its corresponding turning sequence, consisting of several key motion points connected end to end. If the teaching ends at a long straight segment without generating a new significant turning angle, the end pose at the moment of exiting the teaching mode is saved as a virtual key motion point to ensure that the last segment of the path has a clear endpoint and avoid the problem of missing end straight segments.

[0186] Furthermore, any two adjacent key action points and A common target line segment is defined as follows:

[0187]

[0188] in,

[0189]

[0190] , , The correlation coefficient is... These are the coordinates of the point.

[0191] Thus, the entire inspection path is represented as an ordered combination of several continuous target straight lines and the turning actions between them, transforming the subsequent navigation control from "tracking the dense original trajectory" into a deterministic process of "tracking line segments in sequence - executing turning - switching to the next line segment".

[0192] S52: Enhanced calibration inertial positioning is performed based on the drive encoder and inertial measurement unit, the heading angle is integrated to suppress cumulative drift, and a dual square figure-eight trajectory is used to correct the odometer system error;

[0193] During the path reproduction phase, the robot does not rely on SLAM mapping results but instead performs inertial localization based on the drive encoder and inertial measurement unit. The industrial control computer updates the robot's pose at a fixed frequency, and the pose update formula is as follows:

[0194]

[0195]

[0196]

[0197] in,

[0198]

[0199] , Let be the robot's position coordinates at the k-th update time. , Let be the position coordinates of the robot at the (k-1)th update time. Let be the linear displacement increment of the robot in the k-th update cycle. Let be the robot's heading angle at the k-th update time. Let be the robot's heading angle at the (k-1)th update time. This represents the robot's heading angle increment during the k-th update cycle. and These are the displacement increments for the right and left wheels, respectively. This refers to the wheel track.

[0200] To reduce the cumulative drift of the pure odometer, the heading angle of the inertial measurement unit and the odometer heading are further weighted and fused to obtain the fused heading angle:

[0201]

[0202] in, The heading angle is measured by the inertial measurement unit. This is to incorporate weights. This structure enables the robot to continuously obtain relatively stable pose estimates even in cage environments with narrow passages, highly repetitive environments, and difficulty in stable mapping.

[0203] To ensure high-fidelity reproduction of the taught path over long distances, the robot's odometry system error is calibrated before teaching the path. This calibration does not use a standard square test path, but rather a double-square figure-eight test trajectory that better stimulates composite error modes. The endpoint direction error is used as the primary correction, thus more sensitively reflecting the cumulative angular drift during long-distance inspections. This invention classifies system errors into Type A and Type B errors. Type A errors represent a uniform cumulative angular deviation independent of the direction of movement, while Type B errors represent asymmetrical deviations in clockwise and counterclockwise motion.

[0204] For Type A errors caused by inconsistent wheel diameters between the left and right wheels, the following correction formula is used:

[0205]

[0206]

[0207] in, and The corrected wheel diameters for the left and right wheels. and These are the initially measured equivalent wheel diameters of the left and right wheels, respectively. To accumulate directional deviation, This represents the change from a theoretical perspective.

[0208] For Type B errors caused by wheel track error, the following correction formula is used:

[0209]

[0210] in, This is the corrected wheelbase. To initially measure the wheelbase, This represents the imbalance between clockwise and counterclockwise directions. and These represent the cumulative angle changes during clockwise and counterclockwise tests, respectively.

[0211] Through the above calibration, the robot's underlying kinematic parameters were corrected to more closely resemble the actual mechanical state, thereby suppressing the problem of embedding systematic errors into the path model during teaching and learning. In practice, the corrected wheelbase and wheel diameter parameters significantly improved long-distance pose accuracy, and achieved an end-effector relative positioning error of 0.053% in an 85 m cage-raising channel test.

[0212] S53: High-precision path tracking is achieved by synchronously converging lateral and heading deviations through feedback linearization;

[0213] During the path reproduction phase, the industrial control computer does not control the robot solely based on the heading error. Instead, it simultaneously constructs two geometric error quantities: lateral deviation and heading deviation, and achieves synchronous convergence of the two errors through a feedback linearization method.

[0214] For the current target straight line segment The lateral deviation of the robot's current pose point from this line segment is:

[0215]

[0216] in, This represents the robot's current pose coordinates.

[0217] The target heading angle for the target straight segment is:

[0218]

[0219] The robot's current heading deviation is:

[0220]

[0221] in, This represents the robot's current heading angle.

[0222] In the control law design, the lateral deviation and heading deviation are first transformed into second-order convergent error dynamics:

[0223]

[0224]

[0225] in, This indicates lateral deviation error. Indicates the heading deviation error. The control gain is positive. The robot's linear velocity is then generated. and angular velocity control quantity :

[0226] ,

[0227] Then convert it to the speed of the left and right wheels:

[0228] ,

[0229] in, This refers to the distance between the left and right wheels.

[0230] In this way, when the robot tracks each straight line segment, both lateral and heading deviations are compressed simultaneously, instead of just correcting the heading like a conventional PID controller. Therefore, it is more suitable for narrow environments with very little lateral margin, such as multi-level cage passageways.

[0231] Once the robot reaches the end of the current target straight segment, the system calls the steering angle recorded at that key motion point to control the robot to turn in place or make a small-radius turn, aligning its heading with the next target straight segment. Then, it switches to tracking control of the next straight segment. This forms a patrol navigation loop of "tracking the current segment – ​​performing a turn – switching to the next segment" until the entire teaching path is completed.

[0232] S54: Divide obstacles into dynamic obstacles, small static obstacles, and large static obstacles, and implement real-time avoidance, circular detour, and parking waiting strategies respectively;

[0233] During path reconstruction, the left front lidar and right rear lidar continuously scan the surrounding obstacles, and the result of a single scan is represented as:

[0234]

[0235] in, For the first The second scan The distance from each sampling point to the center of the sensor This corresponds to the azimuth angle. The industrial control computer first calculates the diagonal length of the obstacle based on the sampling points at the beginning and end of the same obstacle contour:

[0236]

[0237] , The distance from the first and last sampling points of the same obstacle contour to the center of the sensor. , The azimuth angles corresponding to the first and last sampling points are used to determine the size of the obstacle.

[0238] Furthermore, to distinguish between static and dynamic obstacles, the system combines two adjacent frames of laser scan data with the robot's own mileage displacement. Calculate the two angle quantities separately:

[0239]

[0240]

[0241] in, This is the distance from the reference obstacle point to the center of the sensor in the previous frame scan. This represents the distance from the corresponding obstacle point to the center of the sensor in the current frame scan. This is the azimuth angle of the corresponding obstacle point in the current frame scan.

[0242] If the two are essentially the same, it indicates that the obstacle is stationary relative to the environment and should be classified as a static obstacle. If there is a significant difference, it indicates that the obstacle is moving independently and should be classified as a dynamic obstacle. This classification result is used to switch between different obstacle avoidance strategies. Based on scale and motion attributes, obstacles are divided into three categories: small static obstacles, large static obstacles, and dynamic obstacles.

[0243] When the system detects a dynamic obstacle, the ultrasonic obstacle avoidance sensor provides near-field distance information and performs real-time steering corrections to the robot according to a preset safety threshold. The left and right wheel speed corrections are expressed as follows:

[0244]

[0245] in, The wheel speed differential correction amount generated for obstacle avoidance control. The larger the size, the stronger the robot's steering ability. A proportionality coefficient that is inversely proportional to the distance to the obstacle; the closer the obstacle, the higher the coefficient. The larger the radius, the stronger the robot's steering ability, ensuring rapid avoidance of moving people or suddenly intruding objects. For slow-moving dynamic obstacles, the robot is allowed to deviate slightly from its original path and return after the obstacle is removed; for fast-moving obstacles, it avoids the obstacle with the minimum necessary steering and resumes original path tracking as quickly as possible. Experimental results show that in a two-stage intervention scenario involving a person, the robot can return to its original path within 3.2 seconds after a dynamic pedestrian leaves, with a final residual lateral error of only 0.031 m.

[0246] When the obstacle is determined to be static and satisfies At this time, the system treats it as a small, static obstacle that can be bypassed. Based on the orientation of the obstacle's center relative to the robot's central axis, it automatically chooses to bypass to the left or right, constructing a circular trajectory centered on the obstacle's outer safety boundary. The radius of the arc is:

[0247]

[0248] in, This defines a safe distance between the robot and obstacles. When navigating in a circular path, the left and right wheels satisfy the following conditions:

[0249]

[0250] in, The half-track of the robot, and based on the travel arc length. Real-time calculation of the current turning angle:

[0251]

[0252] During obstacle avoidance, the system continuously updates the diagonal length and relative position of obstacles, thereby correcting the arc radius and turning duration in real time. When an obstacle leaves the perception area or the robot's heading meets the cut-back conditions again, the system exits the arc avoidance and re-cuts back to tracking the target straight line segment. Compared with fixed-radius obstacle avoidance, this arc obstacle avoidance method based on real-time updates of obstacle scale can reduce unnecessary yaw and reduce the adjustment amount required to return to the original path after obstacle avoidance.

[0253] When the obstacle is determined to be static and satisfies At this point, the system treats it as a large, static obstacle that cannot be safely bypassed. The robot does not attempt to force its way around the obstacle, but instead moves at a safe forward distance. When the conditions are met, the robot stops immediately and continuously monitors whether the obstacle has been removed or reclassified as a small static obstacle. If the obstacle's size subsequently decreases to a manageable range, the robot automatically switches to a small static obstacle detour strategy; otherwise, it remains stopped and waits until the obstacle is removed. This strategy prevents the robot from attempting high-risk detours around large obstacles in narrow cage passageways.

[0254] S55: Priority arbitration is performed on the path tracking control output and obstacle avoidance control output. Under normal conditions, path tracking control is in effect; when an obstacle is detected and the obstacle avoidance condition is triggered, obstacle avoidance control takes priority; after the obstacle avoidance is completed, it switches back to path tracking control.

[0255] To enable the robot to smoothly switch between normal inspection, obstacle avoidance, and path return, the industrial control computer prioritizes and arbitrates the path tracking control output and obstacle avoidance control output. The total control input is represented as follows:

[0256]

[0257] in, To provide feedback for linearized path tracking control inputs, This is the obstacle avoidance control parameter. When there are no obstacles, The robot is driven entirely by path tracking control; when an obstacle is detected and obstacle avoidance conditions are triggered, Priority is granted, and path tracking control is temporarily overlaid; after obstacle avoidance is completed... Upon exiting, the system re-enables path tracking control, enabling the robot to re-align with the original taught path and continue executing subsequent inspection segments. Through the aforementioned priority arbitration, the robot establishes a closed-loop autonomous control mode within the poultry house passageway: "autonomous track-following inspection—autonomous obstacle avoidance—autonomous return to track after obstacle clearance—continued inspection."

[0258] S56: Calculate the consistency index between the reproduced path and the taught path to evaluate navigation quality, equipment debugging and acceptance;

[0259] To verify the quality of navigation path reproduction, the industrial control computer evaluates the consistency between the reproduced path and the taught path. The consistency index is defined as:

[0260]

[0261] in, For the teaching path Coordinates of key action points To reproduce the coordinates of the key action points corresponding to the path, This is the total length of the teaching path. The closer the value is to 1, the higher the path reproduction fidelity. This indicator can be used for equipment installation and commissioning, factory calibration, and on-site path update effect acceptance. This invention has been verified in actual cage-raising scenarios, achieving a path reproduction overlap rate of over 98% for both dual-channel and single-channel reciprocating paths, with an average overlap rate exceeding 99% in multiple repeated experiments.

[0262] In summary, the proposed navigation control method organically combines teaching-based path building, enhanced calibration and pose estimation, dual-error feedback linearized trajectory tracking, obstacle classification and processing, and automatic path recovery after obstacle avoidance, thus forming a comprehensive navigation technology solution suitable for narrow, repetitive passages in multi-level cage rearing. Compared to navigation methods that rely on maps and infrastructure, this solution features faster path definition, lower computational burden, less long-distance drift, and stronger trajectory recovery capability after obstacle avoidance, making it more suitable for deployment of poultry egg-collecting robots in real-world cage rearing scenarios.

[0263] S6: After a single inspection channel or a single inspection task is completed, the industrial control computer organizes the inspection results into a structured record at the cage level and uploads it to the cloud server through the communication module. When all tasks are completed, the robot returns to the charging base to dock and recharge.

[0264] To ensure that the aforementioned overall structure, visual inspection method, egg cage matching method, and navigation method form a practical and traceable closed loop for inspection applications, this invention also performs structured processing, local caching, anomaly detection, retransmission and recovery, and task archiving on the counting results during a single inspection task. This guarantees the continuity, integrity, and verifiability of cage-level egg production data. The specific process is as follows:

[0265] The inspection results are generated in a structured manner. During the robot's inspection, identification, matching, and counting processes along the inspection path, the industrial control computer generates cage-level structured records only for results that meet the criteria of "stable QR code identification, valid egg trajectory, unique egg cage ownership, and valid dwell time counting." These structured records include at least the inspection time, cage number, egg quantity, identification confidence level, corresponding trajectory identifier, and evidence frame index or key image information. This transforms the raw image stream and identification process into structured results that can be directly stored, queried, and traced.

[0266] Deduplication and result consolidation within the same cycle. For multiple candidate records corresponding to the same cage location within the same inspection cycle, the industrial control computer performs deduplication and merging based on trajectory identifier, matching establishment time, and counting status. Once a certain type of egg trajectory has been effectively counted, subsequent consecutive frames are only used to supplement evidence images or update status, and are no longer repeatedly accumulated. This avoids the same hatching egg being repeatedly recorded in the same cage location result during the robot's continuous patrol.

[0267] Online uploading and local storage. When the communication link is normal, the communication module uploads the structured records and corresponding evidence data to the cloud management terminal, and marks the local task status as uploaded after the upload is completed; the industrial control computer simultaneously retains the local summary information of this inspection task for subsequent querying, breakpoint recovery, and result verification. This forms an online data flow method of "inspecting, recording, and uploading simultaneously".

[0268] Buffering and retransmission during communication interruptions. When the communication link is temporarily unavailable, upload times out, or data transmission fails, the industrial control computer does not discard the current inspection results. Instead, it writes the unuploaded structured records, key images, and task status into a local cache queue. After the network is restored, the communication module automatically performs retransmission according to the record generation order and removes the cache mark from each successfully retransmitted data. This ensures that inspection data remains continuous and complete even under conditions of fluctuating or temporary network outages in the poultry house wireless network.

[0269] Abnormal Result Detection. For abnormal results encountered during the inspection process, the industrial control computer does not directly output as a valid count, but instead initiates an abnormality detection process. Abnormal situations include at least: QR codes failing to form stable numbers within a preset time window, hatching egg trajectories failing to establish unique cage location assignments, matching relationships repeatedly switching within a short period, images exhibiting significant blurring or occlusion leading to insufficient evidence, and the robot failing to complete effective observation of the current cage location due to obstacle stopping or detouring. For the above situations, the system marks the corresponding cage location as pending verification, requiring rescanning, or a low-confidence result to prevent abnormal frames from directly entering the final counting table.

[0270] Resume inspection and task recovery. When the robot interrupts its current task during inspection due to temporary obstacles, communication failures, manual intervention, or insufficient battery power, the industrial control computer saves information on completed channels, completed cage ranges, unuploaded records, and unfinished inspection sections. Once the interruption is resolved, the robot can resume the task from the most recently unfinished section without having to restart the entire channel inspection. This improves the robot's continuous operation and task recovery capabilities in real-world aquaculture environments.

[0271] Anomaly list output and manual review interface. After a single inspection task is completed, the industrial control computer generates an anomaly list in addition to outputting the normal cage count results. The anomaly list includes at least low-confidence cages, unidentified cages, cages awaiting re-scanning, cages with unstable matching, and cages that have not been confirmed after interruption recovery. Managers can perform manual review or re-inspection based on this anomaly list, thus forming a collaborative working method of "automatic robot inspection and statistics + key manual review".

[0272] Task archiving and closed-loop completion. After a single inspection channel or a single inspection task is completed, the industrial control computer archives the structured record, anomaly list, cache status, and task execution status of this task. After the communication module completes the result synchronization, the robot returns to the charging base for docking and charging. This forms a data closed-loop and anomaly handling process of "task execution - structured recording - anomaly detection - cache retransmission - manual review - result archiving".

[0273] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of the invention.

Claims

1. A control method for a breeding poultry egg-checking robot for multi-layer cage rearing scenarios, based on a breeding poultry egg-checking robot for multi-layer cage rearing scenarios, the breeding poultry egg-checking robot includes a patrol egg-checking robot body (1), a mobile chassis (2), a left front lidar (3), a right rear lidar (4), a left extension bracket (5), a right extension bracket (6), a left upper infrared camera (7), a left lower infrared camera (8), a right upper infrared camera (9), a right lower infrared camera (10), an industrial control computer (11), a chassis drive and CAN control unit (12), a communication module (13), a power supply and battery module (14), a recharge base (15), a cage position QR code label (16), an ultrasonic obstacle avoidance sensor (17), an inertial measurement unit (18), a drive encoder (19), and a touch screen display (20). The left front lidar (3) is located on the left front side of the mobile chassis (2), and the right rear lidar (4) is located on the right rear side of the mobile chassis (2), forming a complementary environmental perception structure. The left extension bracket (5) and the right extension bracket (6) are fixed on the left and right sides of the mobile chassis (2) respectively and extend vertically. The left upper infrared camera (7) and the left lower infrared camera (8) are installed on the left extension bracket (5), and the right upper infrared camera (9) and the right lower infrared camera (10) are installed on the right extension bracket (6). The left upper infrared camera (7) and the right upper infrared camera (9) are installed in a downward-looking manner, and the left lower infrared camera (8) and the right lower infrared camera (10) are installed in an upward-looking manner, forming a composite field of view covering the front edge area of ​​the upper, middle and lower cage positions, the egg trough area and the cage position QR code label (16) area. The cage position QR code label (16) is located at the horizontal center position on the front side of each cage position. The industrial control computer (11) is connected to the left front lidar (3), right rear lidar (4), left upper infrared camera (7), left lower infrared camera (8), right upper infrared camera (9), right lower infrared camera (10), chassis drive and CAN control unit (12), communication module (13), ultrasonic obstacle avoidance sensor (17), inertial measurement unit (18) and drive encoder (19). The recharge base (15) is set at the end of the poultry house, the station area or the preset docking area; the ultrasonic obstacle avoidance sensor (17) is set at the front side or the two front corners of the mobile chassis (2); the inertial measurement unit (18) and the drive encoder (19) are installed inside the mobile chassis (2), wherein the inertial measurement unit (18) is used to output the robot's heading angle, angular velocity or attitude change information, and the drive encoder (19) is used to output the displacement increment of the left and right drive wheels or the travel distance information; The power supply and battery module (14) is used to provide working power for the mobile chassis (2), the left front lidar (3), the right rear lidar (4), the left upper infrared camera (7), the left lower infrared camera (8), the right upper infrared camera (9), the right lower infrared camera (10), the industrial control computer (11) and the communication module (13). Its features are, The control method includes the following steps: S1: The communication module receives the inspection task, the industrial control computer completes the initialization of sensor and chassis parameters, and loads path information, camera calibration parameters and counting judgment parameters; S2: The chassis drive and CAN control unit control the mobile chassis to autonomously inspect along a preset path, collect environmental contour information through the left front lidar and right rear lidar, and simultaneously collect multi-layer cage position images through the left upper infrared camera, left lower infrared camera, right upper infrared camera and right lower infrared camera. S3: The industrial control computer uses a built-in dual-target collaborative detection algorithm for hatching eggs and QR codes to simultaneously detect hatching egg targets and cage location QR code targets, thereby obtaining stable cage location identifiers and hatching egg trajectory information. S4: The industrial control computer uses a built-in precise matching algorithm for egg cages based on time-series trajectory constraints and dynamic neighborhood affiliation. It uses stable cage position identifiers as cage position semantic anchors and egg trajectory information as matching objects to complete the unique affiliation determination between eggs and cage positions. When the stable matching conditions are met, it performs effective counting. S5: The industrial control computer uses a built-in poultry house self-propelled inspection and navigation algorithm based on single-teaching path building, high-precision inertial reproduction and hierarchical obstacle avoidance and return to track. It continuously adjusts the robot's motion posture according to the current position, target path and obstacle status. After the obstacle is removed, the original inspection path is restored. S6: After a single inspection channel or a single inspection task is completed, the industrial control computer organizes the inspection results into a structured record at the cage level and uploads it to the cloud server through the communication module. When all tasks are completed, the robot returns to the charging base to dock and recharge.

2. The control method for the egg-collecting robot for multi-layer cage rearing scenarios according to claim 1, characterized in that, The dual-target collaborative detection algorithm for hatching eggs and QR codes includes the following steps: S31: After distortion correction, temporal alignment and effective field-of-view cropping of the original multi-layer cage position image, shallow edge texture features and deep semantic features are extracted by a shared backbone network. S32: Perform frequency domain feature enhancement processing and illumination sensing parallel fusion on shallow edge texture features and deep semantic features, and input the fused features into the dual-task detection head to output QR code candidate boxes and hatching egg candidate boxes respectively; S33: After performing confidence screening and overlap suppression on the QR code candidate box and the hatching egg candidate box, the hatching egg detection box and the QR code detection box in the current frame are obtained; S34: When the QR code detection frame simultaneously meets the trigger conditions of both global throttling and single trajectory throttling, the QR code image corresponding to the current trajectory is decoded, and the decoding result is written as a stable cage position identifier into the persistent decoding identifier register area, which remains valid throughout the life cycle of the trajectory.

3. The control method for the egg-collecting robot for multi-layer cage rearing scenarios according to claim 2, characterized in that, The frequency domain feature enhancement includes: The feature map is divided into local blocks, and a discrete cosine transform is performed on each local block to obtain the frequency domain feature map. ; in, It is the discrete cosine transform matrix. This is the feature map after local block reorganization. It is the transpose symbol; Adaptive reweighting of different frequency bands in the frequency domain feature map is performed using convolutional layers, nonlinear activation layers, and the sigmoid function to obtain the frequency band weight matrix. The weighted frequency domain features are then inversely transformed back to the spatial domain to form the enhanced spatial features. in, The spatial features are recovered after frequency domain enhancement. For element-wise multiplication; The weighted frequency domain features are inversely transformed back to the spatial domain and then residual-stacked with the original feature map. in, This is the feature map after residual superposition. For learnable balance coefficients, This is the original feature map.

4. The control method for the poultry egg-collecting robot in a multi-layer cage rearing scenario according to claim 3, characterized in that, The parallel fusion of illumination sensing includes: The shallow edge texture features are concatenated with the upsampled deep semantic features and normalized, and then channel threshold denoising and spatial threshold denoising are performed in parallel. The channel threshold denoising is used to adaptively adjust the response intensity of each channel according to the image brightness level, and the spatial threshold denoising is used to suppress spatial noise caused by local reflections and uneven lighting. In the channel threshold denoising branch, global average pooling and global max pooling are used together to generate the brightness perception threshold: in, The brightness perception threshold, It is the Sigmoid activation function. For learnable weights, It is a linear transformation function. This is a global average pooling operation. This is the normalized feature map. This is a global max pooling operation; In the spatial threshold denoising branch, global and local brightness distributions are extracted along the spatial dimension, and a spatially adaptive threshold is generated: in, For spatial adaptive threshold, For convolution operations, For feature splicing operations, These are learnable weights.

5. The control method for the egg-collecting robot for multi-layer cage rearing scenarios according to claim 4, characterized in that, The precise matching algorithm for egg cages based on temporal trajectory constraints and dynamic neighborhood affiliation includes the following steps: S41: For each hatching egg trajectory, calculate the average Euclidean distance between it and each QR code trajectory in the candidate QR code set within the sliding time window; S42: Introduce perspective compensation correction, determine the dynamic matching center by constructing an aspect ratio deviation curve, and define the effective matching range based on the center; S43: Within the effective matching range, determine the optimal matching QR code according to the principle of minimum average Euclidean distance; S44: Only when the optimal matching relationship is maintained within the sliding time window is the matching relationship confirmed as a valid egg cage matching result, and the hatching eggs are assigned to the corresponding cage position number.

6. The control method for the egg-collecting robot for multi-layer cage rearing scenarios according to claim 5, characterized in that, The poultry house self-propelled inspection and navigation algorithm based on single-teaching path construction, high-precision inertial reproduction, and graded obstacle avoidance and return to the track includes the following steps: S51: Manual remote control teaching records the inspection path, extracts key action points where the heading deviation exceeds the threshold, and compresses the continuous trajectory into segmented straight paths and turning sequences; S52: Enhanced calibration inertial positioning is performed based on the drive encoder and inertial measurement unit, the heading angle is integrated to suppress cumulative drift, and a dual square figure-eight trajectory is used to correct the odometer system error; S53: High-precision path tracking is achieved by synchronously converging lateral and heading deviations through feedback linearization; S54: Divide obstacles into dynamic obstacles, small static obstacles, and large static obstacles, and implement real-time avoidance, circular detour, and parking waiting strategies respectively; S55: Priority arbitration is performed on the path tracking control output and obstacle avoidance control output. Under normal conditions, path tracking control is in effect; when an obstacle is detected and the obstacle avoidance condition is triggered, obstacle avoidance control takes priority. After obstacle avoidance is completed, switch back to path tracking control; S56: Calculate the consistency index between the reproduced path and the taught path, which is used to evaluate navigation quality, equipment debugging and acceptance.

7. The control method for the egg-collecting robot for multi-layer cage rearing scenarios according to claim 6, characterized in that, When correcting the odometer system error, corrections are made for Type A errors caused by inconsistent wheel diameters and Type B errors caused by inconsistent wheel track. The correction formula for Type A error is: in, and The corrected wheel diameters for the left and right wheels. and These are the initially measured equivalent wheel diameters of the left and right wheels, respectively. To calculate the cumulative directional deviation, This represents the change from a theoretical perspective. The correction formula for Type B error is: in, This is the corrected wheelbase. To initially measure the wheelbase, This represents the imbalance between clockwise and counterclockwise directions. and These represent the cumulative angle changes during clockwise and counterclockwise tests, respectively.

Images