Control methods and systems for intelligent feeding in ranches

By using an improved YOLOv10n model and camera components to identify feed density and cow behavior, a feeding strategy is generated, which solves the shortcomings of traditional feeding robots in terms of accurate identification and positioning, and realizes personalized feeding and environmental adaptability for dairy cows.

CN121014526BActive Publication Date: 2026-11-17INNER MONGOLIA UNIV OF TECH
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202511138545.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-08-14
Publication Date
2026-11-17
Estimated Expiration
2045-08-14

AI Technical Summary

Technical Problem

Traditional feed-pushing robots lack the ability to accurately identify and locate feed in the feed troughs of dairy cows' feeding areas and the cows themselves, thus failing to meet the personalized feeding needs of dairy cows and making it difficult to adapt to the dynamic changes in complex pasture environments.

Method used

By using an improved YOLOv10n model combined with camera components and sensors, feed density and dairy cow foraging behavior are identified in real time, and corresponding feeding strategies are generated to control the feeding robot to perform precise feeding operations.

Benefits of technology

It improves the ability to accurately identify and locate feed and foraging cows in the feeding troughs of dairy cows, meeting the personalized feeding needs of dairy cows and adapting to dynamic changes in complex pasture environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121014526B_ABST
    Figure CN121014526B_ABST
Patent Text Reader

Abstract

This application discloses a control method and system for intelligent feed pushing in dairy farms. The control method includes: acquiring images of the dairy cow feeding area using a camera component located above the cow feeding area to obtain image data; inputting the image data into a trained image recognition model, and using the image recognition model to identify feed density and cow foraging behavior in the image data, and outputting a first recognition result corresponding to feed density and a second recognition result corresponding to cow foraging behavior; the image recognition model includes an improved YOLOv10n model; generating a corresponding feed pushing strategy based on the first and second recognition results; and controlling a feed pushing robot to perform feed pushing operations according to the feed pushing strategy. Using this application can improve the accurate identification and positioning capabilities of feed in the feed troughs and foraging cows in the dairy cow feeding area, meeting the personalized feeding needs of dairy cows and adapting to dynamic changes in complex pasture environments.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of intelligent feeding technology for ranches, and in particular to a control method and system for intelligent feeding in ranches. Background Technology

[0002] In modern smart dairy farming, precision feeding and feed management directly impact a farm's production efficiency and economic benefits. The rational allocation and full utilization of feed are key to optimizing farming costs, improving cow health, and increasing milk production. However, many farms still rely on manual feeding or fixed-pattern feed delivery. This approach is not only labor-intensive and inefficient but also prone to feed waste and uneven distribution, failing to meet the requirements of refined management in smart farms. In large-scale farming environments, feed delivery involves complex environmental factors and the dynamic behavioral characteristics of foraging cows. Therefore, achieving efficient automated feed delivery and optimized feed utilization is a crucial issue that urgently needs to be addressed to achieve integrated farm management.

[0003] With the rapid development of technologies such as artificial intelligence, the Internet of Things, and edge computing, intelligent feed-pushing robots are gradually becoming an effective means of solving feed management problems in dairy farms. By introducing these advanced technologies, real-time monitoring and precise delivery of feed types, distribution, and demand can be achieved, thereby improving feeding efficiency, optimizing resource allocation, and reducing operating costs. Currently, traditional feed-pushing robots mainly rely on fixed-path navigation or simple sensor information for pushing feed, lacking the ability to accurately identify and locate feed in the feed troughs in the dairy cows' feeding area and the foraging cows. This makes it impossible to meet the personalized feeding needs of dairy cows and adapt to the dynamic changes in complex pasture environments. Summary of the Invention

[0004] The purpose of this application is to provide a control method and system for intelligent feed pushing in a ranch. In a first aspect, this application provides a control method for intelligent feed pushing in a ranch, including:

[0005] The camera unit located above the cow feeding area is used to collect images of the cow feeding area and obtain image data;

[0006] Image data is input into a trained image recognition model, which identifies feed density and cow foraging behavior in the image data and outputs a first recognition result for feed density and a second recognition result for cow foraging behavior. The image recognition model includes an improved YOLOv10n model.

[0007] Based on the first and second identification results, a corresponding material pushing strategy is generated;

[0008] Based on the material pushing strategy, control the material pushing robot to perform the material pushing operation.

[0009] In some embodiments, a corresponding material pushing strategy is generated based on the first identification result and the second identification result, including:

[0010] If the first identification result is sufficient feed, then the first feeding strategy is generated;

[0011] If the first identification result is medium feed, then a second feeding strategy is generated;

[0012] If the first identification result is a small amount of feed and the second identification result is that dairy cows are foraging, then a third feed pushing strategy is generated.

[0013] If the first identification result is a small amount of feed and the second identification result is no dairy cows foraging, then a fourth feed pushing strategy is generated.

[0014] If the first identification result is no feed, then the fifth feeding strategy is generated.

[0015] In some embodiments, the first pushing strategy includes: pushing plate unfolding angle: 30 degrees; pushing robot walking direction: along the parallel direction of the feed trough; operation description: pushing excess feed from the outside of the feed trough back into the feed trough.

[0016] The second feeding strategy includes: the feeding plate unfolding angle: 45 degrees; the feeding robot walking direction: along the parallel direction of the feed trough; operation description: push the feed outside the feed trough back into the feed trough, and adjust the feed distribution to make the feed in the feed trough evenly distributed.

[0017] The third feeding strategy includes: the feed pusher angle is changed from 90 degrees to 30 degrees to collect feed and then unload it, where 90 degrees represents collecting feed and 30 degrees represents unloading feed; operation description: first collect the scattered feed, and then push the feed to the foraging cows.

[0018] The fourth feeding strategy includes: feeding plate unfolding angle: 90 degrees; operation description: clearing away the remaining feed;

[0019] The fifth feeding strategy includes: feeding plate unfolding angle: 90 degrees; operation description: clear debris in the feed trough and send an alarm for insufficient feed to the system.

[0020] In some embodiments, the neck network of the improved YOLOv10n model includes at least a first Zoom_cat splicing module, a second Zoom_cat splicing module, a dynamic scaling module, and an Add module;

[0021] The first Zoom_cat stitching module is connected to the backbone network through a convolutional layer;

[0022] The second Zoom_cat splicing module is connected to the backbone network through a convolutional layer, and is also connected to the first Zoom_cat splicing module through a convolutional layer.

[0023] The dynamic scaling module is connected to the backbone network;

[0024] The Add module is connected to the dynamic scaling module, and is connected to the second Zoom_cat splicing module through a convolutional layer.

[0025] In some embodiments, the training process of the image recognition model includes:

[0026] Obtain an image dataset of the cow feeding area;

[0027] The feed density in the image dataset is classified into several different feed density categories, which are then labeled.

[0028] The cow foraging behaviors in the image dataset were classified to obtain multiple different cow foraging behaviors, and each was labeled.

[0029] An improved YOLOv10n model is constructed based on the backbone network and the neck network, which is at least constructed by the Zoom_cat splicing module, the dynamic scaling module, and the Add module.

[0030] The labeled image dataset is input into the improved YOLOv10n model, and the improved YOLOv10n model is trained until the first recognition result of the improved YOLOv10n model reaches the first predetermined recognition result and the second recognition result reaches the second predetermined recognition result. The improved YOLOv10n model is then used as the image recognition model.

[0031] In some embodiments, inputting the labeled image dataset into the improved YOLOv10n model and training the improved YOLOv10n model includes:

[0032] The labeled image dataset is input into the backbone network;

[0033] The backbone network performs layer-by-layer downsampling on the input image dataset to obtain multiple first feature maps of different layers and sizes.

[0034] The dynamic scaling module of the neck network dynamically enhances part of the first feature map and outputs multiple second feature maps.

[0035] The first Zoom_cat stitching module of the neck network performs the first adaptive max pooling and average pooling adjustment on a portion of the first feature map to obtain the third feature map;

[0036] The second Zoom_cat splicing module of the neck network performs a second adaptive max pooling and average pooling adjustment on some of the first and third feature maps to obtain the fourth feature map.

[0037] The first target feature map is obtained by adding multiple second and fourth feature maps through the Add module of the neck network;

[0038] The second target feature map is obtained by convolution and concatenation of multiple second, third, and fourth feature maps.

[0039] The feature map obtained by downsampling the second target feature map is concatenated with the feature map output by the backbone network, and then a depthwise separable convolution operation is performed with one of the second feature maps to obtain the third target feature map.

[0040] In some embodiments, the first feature maps of multiple different layers and different sizes include at least a first sub-feature map, a second sub-feature map, and a third sub-feature map; the size of the first sub-feature map is larger than the size of the second sub-feature map, and the size of the second sub-feature map is larger than the size of the third sub-feature map.

[0041] Furthermore, the first Zoom_cat concatenation module of the neck network performs the first adaptive max pooling and average pooling adjustment on a portion of the first feature map to obtain the third feature map, including:

[0042] The size of the first sub-feature map is adjusted by the first Zoom_cat stitching module of the neck network so that the size of the first sub-feature map is the same as the size of the second sub-feature map.

[0043] The size of the third sub-feature map is adjusted using the nearest neighbor interpolation method so that the size of the third sub-feature map is the same as that of the second sub-feature map;

[0044] The adjusted first, second, and third sub-feature maps are spliced ​​together by channel dimension to obtain the third feature map.

[0045] In some embodiments, a portion of the first feature map is dynamically enhanced using a dynamic scaling module of the neck network to output multiple second feature maps, including:

[0046] The dynamic scaling module of the neck network is used to enhance the first sub-feature map by utilizing the first, second, and third sub-feature maps to obtain the enhancement result.

[0047] The enhanced result, the original first sub-feature map, the original second sub-feature map, and the original third sub-feature map are respectively used as the second feature map and output.

[0048] In some embodiments, the dynamic scaling module of the neck network enhances the first sub-feature map using a first sub-feature map, a second sub-feature map, and a third sub-feature map to obtain an enhanced result, including:

[0049] The dynamic scaling module of the neck network adjusts the first, second, and third sub-feature maps to the same spatial size.

[0050] The size weights are dynamically generated for the adjusted first sub-feature map, second sub-feature map, and third sub-feature map using a preset network.

[0051] By using dynamically generated size weights, the adjusted first, second, and third sub-feature maps are fused to generate fused information, and the enhanced result includes the fused information.

[0052] Secondly, embodiments of this application provide a control system for intelligent feed pushing in a ranch, comprising:

[0053] The acquisition module is configured to use a camera component located above the cow feeding area to acquire images of the cow feeding area and obtain image data;

[0054] The recognition module is configured to input image data into a trained image recognition model, and to identify feed density and cow foraging behavior in the image data through the image recognition model, and output a first recognition result corresponding to feed density and a second recognition result corresponding to cow foraging behavior; the image recognition model includes an improved YOLOv10n model;

[0055] The generation module is configured to generate a corresponding material pushing strategy based on the first identification result and the second identification result;

[0056] The control module is configured to control the pushing robot to perform pushing operations according to the pushing strategy.

[0057] This application provides a control method for intelligent feed pushing in a dairy farm. First, a camera unit located above the cow feeding area captures images of the cow feeding area, obtaining image data. Then, the image data is input into a trained image recognition model, which identifies feed density and cow foraging behavior in the image data, outputting a first recognition result for feed density and a second recognition result for cow foraging behavior. The image recognition model includes an improved YOLOv10n model. Next, based on the first and second recognition results, a corresponding feed pushing strategy is generated. Finally, based on the feed pushing strategy, a feed pushing robot is controlled to perform the feed pushing operation. Using the technical solution provided in this application improves the accuracy of identifying and locating feed in the feed troughs and foraging cows in the cow feeding area. The generation of corresponding feed pushing strategies based on the recognition results can meet the personalized feeding needs of dairy cows and adapt to dynamic changes in complex pasture environments. Attached Figure Description

[0058] Figure 1 A flowchart illustrating a control method for intelligent feed pushing in a ranch, provided as an embodiment of this application;

[0059] Figure 2 A technical roadmap for a smart feed pushing control system for a ranch, provided in an embodiment of this application;

[0060] Figure 3 A schematic diagram of the network structure of an improved YOLOv10n model provided in this application embodiment;

[0061] Figure 4 This is a schematic diagram of the annotation of an image dataset for feed density provided in an embodiment of this application; wherein, (a) represents no feed, (b) represents a small amount of feed, (c) represents a medium amount of feed, and (d) represents a sufficient amount of feed.

[0062] Figure 5 This is a schematic diagram of the annotation of an image dataset of dairy cow location and behavior provided in an embodiment of this application; wherein, (a) represents grazing (dairy cows are foraging), and (b) represents grazing (dairy cows are foraging) and proximity to the barn (no dairy cows are foraging);

[0063] Figure 6a The F1 curve is a schematic diagram of the training and validation loss curves of the improved Yolov10n model in this application embodiment.

[0064] Figure 6b This is a training result diagram illustrating the training and validation loss curves of the improved Yolov10n model in an embodiment of this application.

[0065] Figure 6cThe accuracy curve is a schematic diagram of the training and validation loss curves of the improved Yolov10n model in this application embodiment.

[0066] Figure 6d The recall curve is a schematic diagram of the training and validation loss curves of the improved Yolov10n model in this application embodiment.

[0067] Figures 7a to 7f This is a schematic diagram illustrating the recognition results of images collected from a pasture using a trained and improved Yolov10n model, as described in an embodiment of this application.

[0068] Figure 8 This is a schematic diagram illustrating the material pushing operation of the pushing robot provided in an embodiment of this application.

[0069] Figure 9 This is a schematic diagram of the structure of a smart feed pushing control system for a ranch, provided in an embodiment of this application. Detailed Implementation

[0070] Various embodiments and features of this application are described herein with reference to the accompanying drawings.

[0071] It should be understood that various modifications can be made to the embodiments described herein. Therefore, the above description should not be considered as limiting, but merely as an example of embodiments. Other modifications within the scope and spirit of this application will be apparent to those skilled in the art.

[0072] The accompanying drawings, which are included in and form part of this specification, illustrate embodiments of the present application and, together with the general description of the present application given above and the detailed description of the embodiments given below, serve to explain the principles of the present application.

[0073] These and other features of this application will become apparent from the following description of preferred forms of embodiments given as non-limiting examples, with reference to the accompanying drawings.

[0074] It should also be understood that although this application has been described with reference to some specific examples, those skilled in the art can certainly implement many other equivalent forms of this application.

[0075] The above and other aspects, features and advantages of this application will become more apparent when taken in conjunction with the accompanying drawings and in view of the following detailed description.

[0076] Specific embodiments of this application are described thereafter with reference to the accompanying drawings; however, it should be understood that the claimed embodiments are merely examples of this application, which can be implemented in various ways. Well-known and / or repeated functions and structures are not described in detail to avoid unnecessary or redundant details that could obscure the application. Therefore, the specific structural and functional details claimed herein are not intended to be limiting, but merely serve as the basis and representative basis for the claims to teach those skilled in the art to use this application in a variety of substantially any suitable detailed structures.

[0077] This specification may use the phrases “in one embodiment,” “in another embodiment,” “in yet another embodiment,” or “in other embodiments,” all of which may refer to one or more of the same or different embodiments according to this application.

[0078] Currently, traditional feed-pushing robots mainly rely on fixed-path navigation or simple sensor information to push feed. They lack the ability to accurately identify and locate feed in the feed troughs in the dairy cows' feeding area and the foraging dairy cows. They cannot meet the personalized feeding needs of dairy cows and are difficult to adapt to the dynamic changes in complex pasture environments.

[0079] To address the aforementioned issues, this application proposes an embedded intelligent feed pushing control method and system for dairy farms based on information fusion. The aim is to achieve high-precision identification and efficient pushing of feed distribution and density in unstructured pasture environments, thereby meeting the personalized feeding needs of dairy cows and adapting to dynamic changes in complex pasture environments.

[0080] The technical solution of this application will be described in detail below with reference to the accompanying drawings.

[0081] See Figure 1 As shown in the figure, an embodiment of this application provides a control method for intelligent feed pushing in a ranch, comprising:

[0082] S10: Use a camera unit located above the cow feeding area to collect images of the cow feeding area and obtain image data;

[0083] S20, input the image data into the trained image recognition model, and use the image recognition model to identify the feed density and the foraging behavior of dairy cows in the image data, and output the first recognition result corresponding to the feed density and the second recognition result corresponding to the foraging behavior of dairy cows; the image recognition model includes the improved YOLOv10n model;

[0084] S30, Generate a corresponding material pushing strategy based on the first identification result and the second identification result;

[0085] S40 controls the pushing robot to perform pushing operations according to the pushing strategy.

[0086] This application provides a control method for intelligent feed pushing in a dairy farm. First, an image acquisition device located above the cow feeding area is used to capture images of the cow feeding area, obtaining image data. Then, the image data is input into a trained image recognition model, which identifies the feed density and cow foraging behavior in the image data and outputs a first recognition result for the feed density and a second recognition result for the cow foraging behavior. The image recognition model includes an improved YOLOv10n model. Then, based on the first and second recognition results, a corresponding feed pushing strategy is generated. Finally, based on the feed pushing strategy, a feed pushing robot is controlled to perform feed pushing operations. This method improves the accuracy of identifying and locating feed in the feed troughs and foraging cows in the cow feeding area. The corresponding feed pushing strategy generated based on the recognition results can meet the personalized feeding needs of cows and adapt to dynamic changes in complex pasture environments.

[0087] In a specific implementation, the cow feeding area can first be divided into multiple zones (e.g., each zone is 2.5 meters apart). The system acquires scene image information through a camera above the cow feeding area. Combined with the powerful detection capabilities and lightweight features of the improved YOLOv10n, it can identify the foraging cows and feed density in each zone. While maintaining high detection accuracy, it significantly improves the system's frame rate and embedded deployment performance, adapting to the needs of resource-constrained edge computing platforms. At the same time, the RTAB-Map (Real-Time Appearance-Based Mapping) algorithm is used to fuse data from LiDAR, IMU, odometry, and depth cameras to construct a high-precision pasture map. Once the map is built, corresponding feeding strategies can be developed based on the feed distribution in each section, providing spatial reference for subsequent feeding operations. In addition, based on preset navigation points, the robot plans the optimal path on the constructed map using a path planning algorithm (Dijkstra's algorithm), ensuring that the robot can efficiently traverse the entire pasture. The feeding control module and the vision recognition module collaborate efficiently through a unified data interface, achieving closed-loop control from target recognition and path generation to operation execution, comprehensively improving the intelligence and automation level of the operation.

[0088] This application aims to solve the key technical problems of existing feeding systems in terms of recognition accuracy, real-time performance, environmental adaptability and control intelligence, and to develop a farm intelligent feeding control method and system with low-cost deployment, high adaptability and high operating efficiency to meet the actual needs of efficient feeding management in future smart farms.

[0089] The intelligent feed pushing control system for ranches disclosed in this application covers the entire process of data acquisition, feed identification, intelligent decision-making, and feed pushing execution. A technical roadmap for an intelligent feed pushing control system for ranches is shown below. Figure 2 As shown. See also Figure 2 First, the information analysis and decision-making module integrates video images, equipment data, and sensor information (such as temperature, humidity, and light intensity) to analyze cattle behavior, feeding conditions, and environmental parameters from multiple dimensions, providing data support for decision-making. Next, the perception and recognition module uses the YOLO target recognition model to accurately extract cattle feeding behavior, activity status, and feed trough changes, optimizing feeding strategies. The information analysis and decision-making module uploads data to the cloud, utilizing distributed storage and analysis to support remote monitoring and intelligent management, ensuring efficient and collaborative task execution. The intelligent control module constructs a refined environmental map based on laser and visual SLAM technologies, and plans the shortest path using the Dijkstra algorithm to achieve precise feed delivery by the feeding robot. Finally, the integration of these modules forms a comprehensive intelligent monitoring system, improving the level of automated farm operations, optimizing cattle feeding and environmental conditions, and promoting refined livestock management.

[0090] The control system for intelligent feed pushing in this application may include a camera component and a sensor component, wherein the camera component may include an overhead camera and a depth camera, and the sensor component may include a lidar, an inertial measurement unit, and an odometer.

[0091] Above-ground cameras: Composed of multiple high-resolution cameras, each installed above the cows' feeding area (10 meters apart, with each camera covering a range of 5 meters to the left and right of the point directly below it), overlooking the entire pasture and used for large-scale monitoring of cow positions and feed trough conditions.

[0092] Depth camera: Acquires depth information of feed troughs and feed, and combined with image recognition technology, more accurately determines feed density and distribution.

[0093] LiDAR (Light Detection and Ranging): Used to acquire two-dimensional or three-dimensional point cloud data of the environment around the pushing robot in real time, helping the pushing robot to detect and avoid obstacles.

[0094] Inertial Measurement Unit (IMU): Provides acceleration and angular velocity information for the pusher robot, used for attitude estimation and motion compensation.

[0095] Odometer: Measures the displacement and speed of the pushing robot using wheel speed sensors to assist in positioning.

[0096] To meet the individualized feeding needs of dairy cows, corresponding feed delivery strategies were developed based on different feed densities and the distribution of foraging dairy cows.

[0097] In some embodiments, a corresponding material pushing strategy is generated based on the first identification result and the second identification result, including:

[0098] If the first identification result is sufficient feed, then the first feeding strategy is generated;

[0099] If the first identification result is medium feed, then a second feeding strategy is generated;

[0100] If the first identification result is a small amount of feed and the second identification result is that dairy cows are foraging, then a third feed pushing strategy is generated.

[0101] If the first identification result is a small amount of feed and the second identification result is no dairy cows foraging, then a fourth feed pushing strategy is generated.

[0102] If the first identification result is no feed, then the fifth feeding strategy is generated.

[0103] In some embodiments, the first feeding strategy includes: the feeding plate unfolding angle: 30 degrees; the feeding robot walking direction: along the parallel direction of the feed trough; operation description: pushing excess feed from the outside of the feed trough back into the feed trough to ensure uniform feed distribution.

[0104] The second feeding strategy includes: the feeding plate unfolding angle is 45 degrees; the feeding robot's walking direction is parallel to the feed trough; the operation description is: push the feed outside the feed trough back into the feed trough and adjust the feed distribution to make the feed in the feed trough evenly distributed and avoid local accumulation.

[0105] The third feeding strategy includes: the feed pusher angle is changed from 90 degrees to 30 degrees to collect feed and then unload it. Here, 90 degrees represents collecting feed and 30 degrees represents unloading feed. Operation description: first collect the scattered feed, and then push the feed to the foraging cows to ensure that the cows have enough feed.

[0106] The fourth feeding strategy includes: the feed plate unfolding angle is 90 degrees; operation description: clear the remaining feed to prepare for the next feeding.

[0107] The fifth feeding strategy includes: feeding plate unfolding angle: 90 degrees; operation description: clean the debris in the feed trough and send an alarm to the system indicating insufficient feed, prompting the staff to replenish the feed in time.

[0108] In some embodiments, see Figure 3 As shown, the neck network of the improved YOLOv10n model includes at least a first Zoom_cat stitching module, a second Zoom_cat stitching module, a dynamic scaling module, and an Add module;

[0109] The first Zoom_cat stitching module is connected to the backbone network through a convolutional layer;

[0110] The second Zoom_cat splicing module is connected to the backbone network through a convolutional layer, and is also connected to the first Zoom_cat splicing module through a convolutional layer.

[0111] The dynamic scaling module is connected to the backbone network;

[0112] The Add module is connected to the dynamic scaling module, and is connected to the second Zoom_cat splicing module through a convolutional layer.

[0113] This application uses an improved YOLOv10n model to train on an image dataset. The improved YOLOv10 model optimizes feature fusion, dynamic scale adjustment, attention mechanisms, and detection algorithms. These improvements enable the model to perform better when handling objects of different scales, complex backgrounds, and occlusion situations, while also improving detection accuracy and robustness.

[0114] The network structure diagram of the improved Yolov10n model is shown below. Figure 3 As shown, the structure of the backbone network is basically the same as before the improvement. The neck network is improved by using the TFE and SSFF modules from ASF-YOLO and a dynamic scaling module, DynamicScalSeq, is added to effectively utilize the relevant information between all pyramid feature maps. The Zoom_cat concatenation module represents the concatenation operation of the three-branch input in the TFE module.

[0115] The main improvements to the Yolov10n model in this application include:

[0116] (1) Feature fusion optimization: The Zoom_cat stitching module replaces the original nn.Upsample and Concat operations. The Zoom_cat stitching module receives three feature maps of different scales (large, medium, and small), adjusts the size of the larger-scale feature map (l) through adaptive max pooling and average pooling to make it the same size as the medium-scale feature map (m), and adjusts the size of the small-scale feature map (s) using the nearest neighbor interpolation method. Then, the three adjusted feature maps are stitched together along the channel dimension to fuse information from different scales.

[0117] Advantages: The Zoom_cat stitching module adopts a more efficient feature fusion strategy, which can better combine feature information of different resolutions, thereby improving the detection performance of objects of different scales.

[0118] (2) Dynamic scale adjustment:

[0119] A Dynamic ScalSeq module has been added. This module dynamically adjusts the scale of features, helping the model better adapt to objects of different sizes. For example, for small objects, the model can dynamically increase the resolution of the feature map to better capture details; for large objects, the model can appropriately decrease the resolution to improve computational efficiency. This dynamic adjustment mechanism can significantly improve the model's detection accuracy for objects of different scales and enhance its generalization ability.

[0120] (3) Introduce the Add module:

[0121] The Add module is a simple module used to add two tensors along a specified dimension, and is often used for the direct fusion of feature maps.

[0122] Advantages: It provides a direct and efficient way to merge features, which helps to simplify the model architecture and accelerate the computation process.

[0123] See also Figure 3As shown, the improved YOLOv10n model in this application takes a 640×640×3 image as input. First, the backbone network downsamples layer by layer according to P1 / 2→P2 / 4→P3 / 8→P4 / 16→P5 / 32 to obtain 320×320×64 feature maps, 160×160×128 feature maps, 80×80×256 feature maps, 40×40×512 feature maps, and 20×20×1024 feature maps. Two convolutional layers (3×3Conv) quickly extract shallow textures. The C2f convolutional layer (C2f-CSP module) enhances gradient flow and improves the representation of small objects while maintaining lightweight through residuals and cross-stage connections. The channel spatial downsampling module (SCDown) replaces stride-2 convolution with learnable downsampling to reduce information loss. SPPF expands the receptive field with multi-scale pooling. The PSA (Local Self-Attention) module suppresses background noise in both channel and spatial dimensions. The improved neck network then proceeds: The convolutional layers (1×1 Conv) in layers 11-12 first unify the P4 / P5 channels. The first Zoom_cat stitching module uses adaptive max pooling and average pooling to align and fuse the high-resolution P3 with the low-resolution P4, solving the feature misalignment problem of traditional Concat stitching. Layers 15-18 again use the second Zoom_cat stitching module to fuse P2 and P3, significantly enhancing the details of small objects. The DynamicScalSeq module in layer 25 performs adaptive enhancement only on P3 (the feature in layer 4): first, P3, P4, and P5 are each adjusted to the same spatial size; then, a small network calculates dynamic weights for these three features at each pixel / channel position; finally, these weights are used to output the enhanced result after fusing the three information paths, along with the original P4 and P5, to the Add module in layer 26. The Add module adds the original P3 to the enhanced result to obtain the enhanced P3. The final output remains a three-way feature map (layers 18 / 21 / 24), but P3 is dynamically enhanced while P4 / P5 remain unchanged for use by subsequent detection heads. The DynamicScalSeq module dynamically generates scale weights for the P3 / P4 / P5 features to achieve content-adaptive fusion, overcoming the shortcomings of fixed weights in dense or scale-changing scenes. The 26th layer Add module (or ASF attention module) further refines the fused features and suppresses redundant background. The final three-way features, namely the refined P3 feature map (80×80×256), and the C2fCIB-enhanced P4 and P5 (40×40×512, 20×20×1024), are fed into the detection head: by default, v10Detect directly regresses bounding boxes and categories in an anchor-free manner; if RTDETRDecoder is enabled, the Transformer decoder is used to improve recall in complex, occluded, or dense target scenes.While maintaining the lightweight and real-time nature of YOLOv10n, the entire design significantly improves the accuracy of small target detection, scale robustness, and background suppression capabilities through three mechanisms: precise alignment by the Zoom_cat stitching module, dynamic weighting by the DynamicScalSeq module, and feature refinement by the Add / ASF module.

[0124] In some embodiments, the training process of the image recognition model includes:

[0125] Obtain an image dataset of the cow feeding area;

[0126] The feed density in the image dataset is classified into several different feed density categories, which are then labeled.

[0127] The cow foraging behaviors in the image dataset were classified to obtain multiple different cow foraging behaviors, and each was labeled.

[0128] An improved YOLOv10n model is constructed based on the backbone network and the neck network, which is at least constructed by the Zoom_cat splicing module, the dynamic scaling module, and the Add module.

[0129] The labeled image dataset is input into the improved YOLOv10n model, and the improved YOLOv10n model is trained until the first recognition result of the improved YOLOv10n model reaches the first predetermined recognition result and the second recognition result reaches the second predetermined recognition result. The improved YOLOv10n model is then used as the image recognition model.

[0130] Specifically, the first step is to obtain an image dataset of the cows' feeding area.

[0131] Then, the feed density in the image dataset was classified into multiple different feed density categories, which were then labeled. Specifically, using image data captured by a camera above the cows' feeding area, the feed was classified into four categories according to different density distributions, and the labelimg tool was used to label them. The dataset labeling method is as follows: Figure 4 As shown, (a) represents no feed, (b) represents a small amount of feed, (c) represents a medium amount of feed, and (d) represents a sufficient amount of feed; the calibration standards for the feed density dataset are shown in Table 1.

[0132] The annotation method for the feed density image dataset is as follows: The feed trough in the cow feeding area below the camera is divided into four regions, each 2.5 meters wide. The feed in each region is labeled according to the image dataset labeling standard. The feed density is divided into four categories: No forage, Little forage, Medium forage, and Rich in forage. These four categories correspond to different feeding strategies.

[0133] Table 1: Calibration Standards for Feed Density Datasets

[0134] Target number Target Category Target definition Tag information 0 feedless There was no obvious feed in the feed trough, and the feed height was less than 1 cm. No forage 1 Small amount of feed Feed was present in some areas of the feed trough, with a feed height of 2-5cm. Little forage 2 Medium feed There was obvious feed in the feed trough, with a feed height of 5-10cm. Medium forage 3 Sufficient feed There was obvious feed in the feed trough, and the feed height was greater than 10cm. Richin forage

[0135] The image dataset was categorized based on cow foraging behavior, resulting in multiple distinct foraging behaviors, which were then labeled. Specifically, using image data captured by cameras above the cow feeding area, two foraging behaviors were manually selected: grazing (cows are foraging) and approaching the pen (no cows are foraging). These behaviors were then labeled using the labelimg tool. The labeling method for the cow behavior image dataset is as follows: Figure 5 As shown, (a) represents feeding (with cows foraging), and (b) represents feeding (with cows foraging) and proximity to the barn (without cows foraging); the calibration criteria for the cow behavior dataset are shown in Table 2.

[0136] Table 2: Standardization criteria for dairy cow behavior datasets

[0137]

[0138]

[0139] In some embodiments, inputting the labeled image dataset into the improved YOLOv10n model and training the improved YOLOv10n model includes:

[0140] The labeled image dataset is input into the backbone network;

[0141] The backbone network performs layer-by-layer downsampling on the input image dataset to obtain multiple first feature maps of different layers and sizes.

[0142] The dynamic scaling module of the neck network dynamically enhances part of the first feature map and outputs multiple second feature maps.

[0143] The first Zoom_cat stitching module of the neck network performs the first adaptive max pooling and average pooling adjustment on a portion of the first feature map to obtain the third feature map;

[0144] The second Zoom_cat splicing module of the neck network performs a second adaptive max pooling and average pooling adjustment on some of the first and third feature maps to obtain the fourth feature map.

[0145] The first target feature map is obtained by adding multiple second and fourth feature maps through the Add module of the neck network;

[0146] The second target feature map is obtained by convolution and concatenation of multiple second, third, and fourth feature maps.

[0147] The feature map obtained by downsampling the second target feature map is concatenated with the feature map output by the backbone network, and then a depthwise separable convolution operation is performed with one of the second feature maps to obtain the third target feature map.

[0148] For details, see Figure 3 As shown, the input image dataset is downsampled layer by layer through the backbone network to obtain multiple first feature maps of different layers and sizes. Optionally, the backbone network is used to downsample layer by layer in the order of P1 / 2→P2 / 4→P3 / 8→P4 / 16→P5 / 32 to obtain 320×320×64 feature maps, 160×160×128 feature maps, 80×80×256 feature maps, 40×40×512 feature maps, and 20×20×1024 feature maps.

[0149] Then, the dynamic scaling module of the neck network dynamically enhances some of the first feature maps, outputting multiple second feature maps. Optionally, the dynamic scaling module only performs content-adaptive enhancement on P3 (the 4th layer feature): first, P3, P4, and P5 are each adjusted to the same spatial size; then, a small network calculates dynamic weights for these three features at each pixel / channel position; finally, these weights are used to fuse the enhanced result of the three information paths, along with the original P4 and P5, and output to the Add module of layer 26. The Add module adds the original P3 to the enhanced result to obtain the enhanced P3. The final output is still three features (layers 18 / 21 / 24), but P3 is dynamically enhanced while P4 and P5 remain unchanged, and is output to the Add module without any additional processing.

[0150] Then, the first Zoom_cat stitching module of the neck network is used to perform the first adaptive max pooling and average pooling adjustment on part of the first feature map to obtain the third feature map;

[0151] Then, the second Zoom_cat splicing module of the neck network performs a second adaptive max pooling and average pooling adjustment on some of the first and second feature maps to obtain the fourth feature map.

[0152] Then, the Add module of the neck network is used to add multiple second and fourth feature maps to obtain the first target feature map (the refined P3 feature map);

[0153] Then, the multiple second, third, and fourth feature maps are convolved and concatenated to obtain the second target feature map (the depth-separable enhanced P4 feature map);

[0154] Then, the feature map obtained by downsampling the second target feature map is concatenated with the feature map output by the backbone network, and then a depthwise separable convolution operation is performed with one of the second feature maps to obtain the third target feature map (the depthwise separable enhanced P5 feature map).

[0155] In some embodiments, the first feature maps of multiple different layers and different sizes include at least a first sub-feature map, a second sub-feature map, and a third sub-feature map; the size of the first sub-feature map (e.g., P3) is larger than the size of the second sub-feature map (e.g., P4), and the size of the second sub-feature map is larger than the size of the third sub-feature map (e.g., P5);

[0156] Furthermore, the first Zoom_cat concatenation module of the neck network performs the first adaptive max pooling and average pooling adjustment on a portion of the first feature map to obtain the third feature map, including:

[0157] The size of the first sub-feature map (e.g., P3) is adjusted by the first Zoom_cat stitching module of the neck network so that the size of the first sub-feature map (e.g., P3) is the same as the size of the second sub-feature map (e.g., P4);

[0158] The size of the third sub-feature map (e.g., P5) is adjusted using the nearest neighbor interpolation method so that the size of the third sub-feature map (e.g., P5) is the same as the size of the second sub-feature map (e.g., P4).

[0159] The adjusted first, second, and third sub-feature maps are spliced ​​together by channel dimension to obtain the third feature map.

[0160] That is, the convolutional layer (1×1Conv) of layers 11-12 first unifies the P4 / P5 channels. The first Zoom_cat stitching module uses adaptive max pooling and average pooling to adjust and merge the high-resolution P3 with the low-resolution P4, thus solving the problem of feature misalignment in traditional Concat (stitching).

[0161] Then, the third feature map is output to the second Zoom_cat splicing module through C2f convolutional layers and convolutional layer operations.

[0162] The second Zoom_cat concatenation module uses the same method to perform a second adaptive max pooling and average pooling adjustment on the first feature map (such as P2, P3) and the third feature map to obtain the fourth feature map; the fourth feature map is fed into the Add module after passing through the C2f convolutional layer.

[0163] In some embodiments, a portion of the first feature map is dynamically enhanced using a dynamic scaling module of the neck network to output multiple second feature maps, including:

[0164] The dynamic scaling module of the neck network enhances the first sub-feature map (e.g., P3) using the first sub-feature map (e.g., P3), the second sub-feature map (e.g., P4), and the third sub-feature map (e.g., P5) to obtain the enhancement result.

[0165] The enhanced result, the original first sub-feature map, the original second sub-feature map, and the original third sub-feature map are respectively used as the second feature map and output.

[0166] Specifically, the dynamic scaling module outputs multiple second feature maps, which include the enhanced result of P3 and the original P3, P4, and P5.

[0167] In some embodiments, the dynamic scaling module of the neck network enhances the first sub-feature map using a first sub-feature map, a second sub-feature map, and a third sub-feature map to obtain an enhanced result, including:

[0168] The dynamic scaling module of the neck network adjusts the first sub-feature map (e.g., P3), the second sub-feature map (e.g., P4), and the third sub-feature map (e.g., P5) to the same spatial size.

[0169] The size weights are dynamically generated for the adjusted first sub-feature map, second sub-feature map, and third sub-feature map using a preset network.

[0170] By using dynamically generated size weights, the adjusted first sub-feature map, second sub-feature map, and third sub-feature map are fused to generate fused information, and the enhancement result includes the fused information.

[0171] Specifically, P3, P4, and P5 are each adjusted to the same spatial size; then, a small network is used to calculate dynamic weights for these three features at each pixel / channel position; using the dynamically generated size weights, the adjusted first sub-feature map, second sub-feature map, and third sub-feature map are fused to generate fused information, i.e., the enhanced result of P3.

[0172] Specifically, the Add module adds the original P3 to the enhancement result to obtain the enhanced P3. Then, it adds the enhanced P3, the original P4, P5, the fourth feature map obtained by the second Zoom_cat concatenation module after passing through the C2f convolutional layer to obtain the first target feature map (refined P3 feature map).

[0173] Then, the fourth feature map obtained by the second Zoom_cat concatenation module is concatenated with the third feature map obtained by the first Zoom_cat concatenation module after passing through a C2f convolutional layer and a convolutional layer. After passing through a C2f convolutional layer, the second target feature map (the P4 feature map after depth separability enhancement) is obtained.

[0174] Then, the feature map obtained by downsampling the second target feature map (the depth-separable enhanced P4 feature map) through the channel-spatial downsampling module is concatenated with the feature map output by the PSA module of the backbone network through the convolutional layer. The resulting feature map and the second feature map (P3) are then processed through the depth-separable convolutional layer to obtain the third target feature map (the depth-separable enhanced P5 feature map).

[0175] Using the improved YOLOv10n model algorithm, we accurately detected and identified foraging cow behavior and feed density in dairy farms. Regarding cow behavior identification, the YOLOv10n model achieved a recognition accuracy of 96.4% for "Eating" behavior and 77.1% for "Approaching" behavior, both meeting the requirements for accurate identification of foraging cow behavior. In terms of feed density identification, the YOLOv10n model achieved an accuracy exceeding 90% for four different feed densities, providing a solid data foundation for automated feed feeding operations. The experimental results of the original YOLOv10n model are shown in Table 3. The training and validation loss curves of the improved YOLOv10n model are detailed in [link to table]. Figures 6a to 6d , Figure 6a This is the F1 curve. Figure 6b The image shows the training results. Here, `train / box_loss` represents the bounding box loss during training, `train / cls_loss` represents the classification loss during training, `train / dfl_loss` represents the distribution focus loss during training, `metrics / precision(B)` represents the precision metric, `metrics / recall(B)` represents the recall metric, `val / box_loss` represents the bounding box loss on the validation set, `val / cls_loss` represents the classification loss on the validation set, `val / dfl_loss` represents the distribution focus loss on the validation set, `metrics / mAP50(B)` represents the mean precision metric, and `metrics / mAP50-95(B)` represents the average mean precision at different IoU thresholds (from 0.5 to 0.95, with a step size of 0.05). Figure 6c For the accuracy curve, Figure 6d The recall rate curve is shown in Table 4. Specific experimental results are summarized in Table 4. These results demonstrate that the improved YOLOv10n model algorithm can effectively support key aspects of automated feed management in smart animal husbandry.

[0176] Table 3: Training Results of the Original Yolov10n Model

[0177]

[0178]

[0179] Table 4: Training Results of the Improved Yolov10n Model

[0180] Class Precision / % Recall / % mAP50 / % mAP 50-95% all 0.924 0.876 0.911 0.654 Eating 0.964 0.958 0.978 0.547 approach 0.771 0.408 0.535 0.246 No forage 0.993 0.988 0.995 0.768 Little forage 0.98 0.942 0.989 0.783 Medium forage 0.925 0.964 0.978 0.778 Richin forage 0.908 0.997 0.99 0.8

[0181] Based on the training results, the improved YOLOv10n model exhibited good recognition performance at different confidence levels: the F1-Confidence Curve shows that the model achieves a high F1 score at low to medium confidence levels, indicating that the model performs well in balancing precision and recall, especially in the "Rich in forage" and "No forage" categories. The evaluation metrics in the training results graph show that mAP (mean precision) and precision increase with the number of training epochs, indicating that the model's prediction accuracy and recall are improving. The Precision-Confidence Curve shows that the model maintains extremely high precision at high confidence levels, which is crucial for ensuring the accuracy of the feeding operation. The Recall-Confidence Curve shows that the model has a high recall rate at low confidence levels, helping to ensure that no areas requiring feeding are missed.

[0182] By employing an improved YOLOv10n algorithm and deep learning technology, high-precision recognition of cow behavior and feed status in a pasture environment was achieved, demonstrating strong environmental adaptability and high real-time performance. This is crucial for the accurate task execution of automated feed-pushing robot systems in changing pasture environments. The advantages of the improved YOLOv10n model lie in its high-precision positioning capability, strong environmental adaptability, and high real-time performance, making it highly suitable for automated feed management in smart farms. However, to further improve model performance, future work can focus on optimizing the recognition accuracy of the "approach" category and enhancing the model's continuous perception and multi-frame fusion capabilities in dynamic environments. This will ensure the model's stability and accuracy even when cow and feed status changes in real time on the pasture. These improvements can further enhance the efficiency and reliability of automated feed-pushing systems, meeting the practical needs of smart farms for refined management.

[0183] The improved YOLOv10n model demonstrates significant advantages across several key performance indicators, with overall precision and recall increasing to 92.4% and 87.6%, respectively, while mAP50 and mAP50-95 also show improvement. Particularly noteworthy is its accuracy in recognizing "Eating" behavior, reaching 96.4%, while the accuracy for "approach" behavior recognition has significantly improved to 77.1%. Furthermore, recognition performance across all feed density categories has been enhanced, with recall rates for "Medium forage" and "Rich in forage" categories increasing to 96.4% and 99.7%, respectively, and improvements also observed in mAP50 and mAP50-95. These improvements indicate that the improved YOLOv10n model is more accurate and reliable in automated feed management for smart livestock farming, and can more effectively support the decision-making and operation of automated feeding robot systems.

[0184] A detailed analysis was performed on six images of cows feeding areas in a dairy farm using a fully trained and improved YOLOv10n model. The improved YOLOv10n model performed excellently in identifying feed density distribution and cow foraging behavior in each area. Specific prediction results are shown below. Figures 7a to 7f .

[0185] Figure 7a The data showed the identification of 4 "Eating" behaviors, 2 "Medium forage" behaviors, and 4 "Little forage" behaviors, with a processing time of 97.5 milliseconds.

[0186] Figure 7b The model identified one instance of "Medium forage" and four instances of "Little forage" in just 6.0 milliseconds.

[0187] Figure 7c The model identified four instances of "No forage," all within 6.0 milliseconds.

[0188] Figure 7d It demonstrated the recognition of "Eating" once, "Medium forage" four times and "Little forage" once, in 7.0 milliseconds.

[0189] Figure 7e In the study, the model identified one “approach” behavior, four “Medium forage” behaviors, and one “Rich in forage” behavior, with a processing time of 6.1 milliseconds.

[0190] Figure 7f The system displayed 7 instances of "Eating", 1 instance of "approach", 2 instances of "Medium forage" and 4 instances of "Rich in forage", with a processing time of 6.0 milliseconds.

[0191] The overall detection process is very fast, with a preprocessing time of 2.3 milliseconds, an inference time of 21.4 milliseconds, and a postprocessing time of 8.9 milliseconds per image, demonstrating the efficiency and accuracy of the improved YOLOv10n model in practical applications. These results clearly demonstrate the key role of the improved YOLOv10n model algorithm in automated feed management in smart animal husbandry, effectively supporting automated feeding robot systems to accurately perform tasks in changing pasture environments and meeting the practical needs of smart animal husbandry for refined management.

[0192] In some embodiments, the specific steps for controlling the pushing robot to perform the pushing operation according to the pushing strategy can be found in [reference needed]. Figure 8 As shown, the intelligent feed-pushing control system of this ranch utilizes 5G communication technology for remote control and data transmission. The process begins with system startup and self-check, verifying machine malfunctions and load capacity. Subsequently, the control terminal issues a feeding task, including information such as feed distribution density and target location. The feeding robot receives the task via 5G communication, and data transmission is encrypted to ensure security. Next, the system optimizes the path and automatically avoids obstacles, navigating the feeding robot to the target feeding area. After confirming the feeding task, the feeding robot performs the feeding operation, adjusting the unfolding angle of the feeding plate, feeding accuracy, and stability. After the task is completed, the system checks the feeding task. If the task is incomplete, the robot returns to the standby area and restarts the process; if the task is completed, the process ends. The entire process demonstrates the intelligence and efficiency of the automated feeding system, suitable for the refined management of smart animal husbandry.

[0193] In some embodiments, controlling the pushing robot to perform pushing operations includes: using the global positioning algorithm of RTAB-Map and the local positioning algorithm of AMCL to fuse the data collected by the camera component and the sensor component to generate a predetermined pose; and controlling the pushing robot to move (walk) and push materials in the predetermined pose.

[0194] Specifically, data fusion and map construction: The RTAB-Map (Real-Time Appearance-Based Mapping) algorithm is used to fuse data from LiDAR, IMU, odometry, and depth cameras to construct a high-precision pasture map. This algorithm can update the map in real time to adapt to changes in the pasture environment, such as the movement of feed and the movement of cows.

[0195] RTAB-Map (Real-Time Appearance-Based Mapping) is a real-time mapping algorithm based on appearance, widely used in robot navigation, environmental perception, and 3D reconstruction. Its core idea is to identify the environment through visual features (such as SIFT, SURF, or ORB) and combine this with multi-sensor data for map construction and localization. RTAB-Map supports data fusion from multiple sensors, including RGB-D cameras, LiDAR, and IMUs, and can generate dense 3D point cloud maps or 2D raster maps. It also optimizes mapping through a loop closure detection mechanism. Figure 1 Consistency. This algorithm features high accuracy, real-time performance, and flexibility, making it suitable for dynamic environments and complex scenarios.

[0196] To address the need for pasture-pushing robots to perceive their surroundings in real time during operation, this application optimizes the RTAB-Map algorithm. Specifically, during mapping, the Cartographer algorithm processes laser, IMU, and odometry data to generate mileage estimates, providing positioning compensation for RTAB-Map. RTAB-Map fuses depth camera point cloud and laser data, corrects global pose through visual loop closure detection, and combines 2D localization to construct a dual-modal environmental representation including a 2D grid map and a 3D point cloud, achieving a complete description of the simulated pasture environment and material morphology. A framework for the robot's navigation algorithm is built based on this fused data. During navigation, RTAB-Map's global localization (low-frequency optimization) and AMCL's local localization (high-frequency update) are combined. Robot_pose_ekf fuses IMU and odometry data to generate a smoother and more robust pose. Dijkstra's algorithm is used for global path planning, combined with the TEB (Timed Elastic Band) algorithm for local path planning.

[0197] By combining data from wheeled odometry, IMU sensors, and LiDAR, the robot's pose estimation accuracy and environmental perception capabilities during material pushing operations are improved. During pushing, wheeled odometry data provides short-term displacement change information, IMU data assists in correcting posture deviations, and LiDAR scans the environment in real time to construct a two-dimensional occupancy grid map and provide obstacle detection capabilities. The improved RTAB-Map algorithm enhances the robot's environmental perception along the pushing path through multi-source data fusion, effectively reducing pose drift caused by uneven ground, obstacles, and other factors, thereby ensuring the continuity and accuracy of the robot's material pushing operations.

[0198] Navigation and Path Planning: Based on preset navigation points, the pusher robot uses the Adaptive Monte Carlo Localization (AMCL) algorithm to achieve precise positioning on a 2D map. A path planning algorithm (Dijkstra's algorithm) plans the optimal path on the constructed map, ensuring the pusher robot can efficiently traverse the entire pasture. Furthermore, the Temporal Elastic Band (Teb) algorithm ensures the pusher robot avoids dynamic obstacles during navigation. During path planning, real-time sensor data is used to dynamically adjust the path and avoid obstacles (such as cows, feed piles, etc.).

[0199] AMCL (Adaptive Monte Carlo Localization) is a probabilistic localization algorithm based on particle filtering, widely used for autonomous localization of mobile robots in known maps. It maintains a set of particles (each representing a possible robot pose) and iteratively updates the pose distribution by combining sensor data and a motion model. The algorithm first randomly initializes particles based on pose uncertainty and propagates the particle positions through the motion model as the robot moves. Then, it calculates the weight of each particle based on sensor data (such as LiDAR), reflecting its degree of matching with the actual environment. Resampling retains high-weight particles and discards low-weight particles to enhance the estimation of the true pose. AMCL can also dynamically adjust the number of particles according to localization uncertainty, improving efficiency while ensuring accuracy. Finally, it outputs the weighted average of the particles as an estimate of the robot's current pose. This method is robust and can achieve efficient real-time localization in dynamic environments.

[0200] Dijkstra's algorithm is a classic single-source shortest path algorithm. In the research of pasture feeding robots, it can be used to efficiently plan the optimal path from the starting point to the target point, improving the efficiency and accuracy of task execution. The pasture environment can be modeled as a weighted graph, where nodes represent key points or working areas of the feeding channel, and edge weights ω(u,v) represent the path cost (such as distance or energy consumption) from node u to node v. The algorithm initializes the distance from the starting point to 0 (i.e., d(s) = 0) and other nodes to infinity (i.e., d(v) = ∞), then progressively selects the node u with the current minimum distance to expand, and updates the shortest distance d(v) of adjacent nodes:

[0201] d(v)=min(d(v),d(u)+ω(u,v))

[0202] This process is repeated until all nodes have been visited or the path to the target node is determined. During the operation of a pasture-pushing robot, Dijkstra's algorithm can be used to plan the shortest path from the charging station to the feeding area, ensuring the lowest possible path cost. Furthermore, when dynamic obstacles appear, real-time path adjustments can be made through recalculation to adapt to complex environments. Finally, combining this with multi-objective path optimization algorithms can also help the robot complete multi-point feeding tasks, thereby achieving efficient feeding management and path planning.

[0203] The Timed Elastic Band (TEB) algorithm is a highly efficient local path planning algorithm, particularly suitable for robot navigation in dynamic environments. In the operation of a pasture-pushing robot, the TEB algorithm optimizes the temporal and spatial distribution of the path to plan a smooth and dynamically obstacle-avoiding trajectory. Based on optimization methods, the algorithm represents the path as a series of discrete points with time constraints, comprehensively considering obstacle constraints, the robot's motion model, and time costs to generate a feasible path that meets practical requirements. Using the TEB algorithm, the pasture-pushing robot can adjust its pushing path in real time, ensuring efficient completion of the pushing task in dynamic environments (such as the random movement of livestock or equipment), while reducing path length and travel time.

[0204] The beneficial effects of the technical solutions provided in this application include at least the following:

[0205] High adaptability: The vision acquisition module can acquire two-dimensional and three-dimensional information of feed, and has good adaptability to complex shapes and irregular feeds. It can formulate corresponding feeding strategies according to the density distribution characteristics of feeds, and can be widely used in a variety of production scenarios.

[0206] High-precision positioning: The RTAB-Map algorithm achieves high-precision positioning of feed by fusing data from multiple sensors and combining it with the accurate identification of the improved YOLOv10 algorithm, thus meeting the needs of high-precision production.

[0207] High real-time performance: Employing a high-performance embedded processor and optimized algorithms, it can complete image acquisition, processing, and control command generation in a short time, meeting the real-time requirements of high-speed production scenarios.

[0208] Strong environmental adaptability: The system can adapt to different lighting conditions and complex backgrounds, and has strong environmental adaptability.

[0209] The key technical points of the technical solution of this application include at least:

[0210] Multi-zone identification and detection strategy: This application divides the feed troughs in the pasture into multiple zones using multiple cameras, enabling precise identification and detection of cow behavior and feed density in each zone. This strategy allows the system to perform differentiated feed pushing operations in different zones based on specific circumstances, thereby improving the precision of feed management.

[0211] Application of the YOLOv10n algorithm: The improved YOLOv10 algorithm is used to quickly and accurately identify feed density and distribution. Its high accuracy and real-time performance make it an ideal choice for this system to identify dairy cow behavior and feed status.

[0212] Multi-sensor data fusion and RTAB-Map navigation: By fusing data from multiple sensors such as LiDAR, IMU, odometry, and depth cameras, and combining this with the RTAB-Map algorithm, high-precision positioning and navigation of the robot are achieved. This technology ensures that the pushing robot can accurately reach the designated location and perform its pushing task in complex pasture environments.

[0213] High-performance embedded processor and optimized algorithms: A high-performance embedded processor is employed, and optimized algorithms are developed to ensure high real-time performance of the system. This optimization not only improves the system's response speed but also guarantees stable operation on resource-constrained embedded devices.

[0214] Intelligent feed delivery strategy: Based on the identification and detection results, the system can automatically generate and execute corresponding feed delivery strategies. This intelligent strategy can dynamically adjust the feed delivery operation according to feed density and dairy cow behavior, achieving precise feed delivery, reducing waste, and improving feed utilization.

[0215] Furthermore, this application improves the accuracy of identifying and locating feed in the feed troughs and foraging cows in the cow feeding area by using a multi-region identification and detection strategy, an improved YOLOv10n model to identify feed density and cow behavior, and an intelligent feed pushing strategy. Based on the identification results, a corresponding feed pushing strategy is generated, which can meet the personalized feeding needs of cows and adapt to the dynamic changes in complex pasture environments.

[0216] Specifically, the multi-area identification and detection strategy is a method that uses multiple cameras to divide the feed troughs in a pasture into multiple areas, enabling accurate identification and detection of cow behavior and feed density in each area.

[0217] Application of the improved YOLOv10n algorithm in feed density identification:

[0218] (1) Feature fusion optimization:

[0219] The Zoom_cat module is used to replace the traditional nn.Upsample and Concat operations, enabling efficient fusion of feature maps at different scales (large, medium, and small). The feature map size is adjusted by adaptive max pooling and average pooling, and the size is matched using the nearest neighbor interpolation method. Then, the maps are stitched along the channel dimension to integrate multi-scale information.

[0220] This improved feature fusion strategy can more effectively combine feature information at different resolutions, thereby improving the detection performance of objects at different scales.

[0221] (2) Dynamic scale adjustment:

[0222] The DynamicScalSeq module is introduced to enable dynamic adjustment of the feature map scale.

[0223] This module can dynamically adjust the feature map resolution based on the object size, optimizing detail capture for small objects and computational efficiency for large objects. This dynamic adjustment mechanism significantly improves the model's detection accuracy for objects of different scales and enhances its generalization ability.

[0224] (3) Introduction of the Add module:

[0225] The Add module allows for the direct fusion of feature maps by adding two tensors along a specified dimension. It provides a direct and efficient method for feature merging, helping to simplify model architecture and accelerate the computation process.

[0226] These improvements enable the YOLOv10n algorithm to exhibit higher accuracy and efficiency in feed density identification, making it suitable for key aspects of automated feed management in smart animal husbandry.

[0227] Intelligent feed pushing strategy: A method that automatically generates and executes corresponding feed pushing strategies based on recognition and detection results. It can dynamically adjust the feed pushing operation according to feed density and dairy cow behavior, so as to achieve precise feed delivery, reduce waste, and improve feed utilization.

[0228] Based on the same inventive concept, see [link to inventive concept] Figure 9 As shown in the figure, this application provides a control system for intelligent feeding in a ranch, including:

[0229] The acquisition module is configured to use a camera component located above the cow feeding area to acquire images of the cow feeding area and obtain image data;

[0230] The recognition module is configured to input image data into a trained image recognition model, and to identify feed density and cow foraging behavior in the image data through the image recognition model, and output a first recognition result corresponding to feed density and a second recognition result corresponding to cow foraging behavior; the image recognition model includes an improved YOLOv10n model;

[0231] The generation module is configured to generate a corresponding material pushing strategy based on the first identification result and the second identification result;

[0232] The control module is configured to control the pushing robot to perform pushing operations according to the pushing strategy.

[0233] The intelligent feed pushing control system for dairy farms provided in this application embodiment, by setting up a data acquisition module, an identification module, a generation module, and a control module, can improve the accuracy of identifying and locating feed in the feed troughs in the dairy cow feeding area and the foraging dairy cows. Based on the identification results, it generates corresponding feed pushing strategies, which can meet the personalized feeding needs of dairy cows and adapt to the dynamic changes in complex pasture environments.

[0234] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A control method for intelligent feeding in a ranch, characterized in that, include: The camera unit located above the cow feeding area is used to collect images of the cow feeding area and obtain image data; The image data is input into a trained image recognition model, and the image recognition model identifies feed density and cow foraging behavior in the image data, and outputs a first recognition result corresponding to feed density and a second recognition result corresponding to cow foraging behavior; the image recognition model includes an improved YOLOv10n model; Based on the first and second identification results, a corresponding material pushing strategy is generated; According to the aforementioned pushing strategy, control the pushing robot to perform pushing operations; Based on the first and second identification results, a corresponding material pushing strategy is generated, including: If the first identification result is sufficient feed, then the first feeding strategy is generated; If the first identification result is medium feed, then a second feeding strategy is generated; If the first identification result is a small amount of feed and the second identification result is that dairy cows are foraging, then a third feed pushing strategy is generated. If the first identification result is a small amount of feed and the second identification result is no dairy cows foraging, then a fourth feed pushing strategy is generated. If the first identification result is no feed, then the fifth feeding strategy is generated; The first feeding strategy includes: pushing plate unfolding angle: 30 degrees; pushing robot walking direction: parallel to the feed trough; operation description: pushing excess feed from the outside of the feed trough back into the feed trough. The second feeding strategy includes: the feeding plate unfolding angle: 45 degrees; the feeding robot walking direction: along the parallel direction of the feed trough; operation description: push the feed outside the feed trough back into the feed trough, and adjust the feed distribution to make the feed in the feed trough evenly distributed. The third feeding strategy includes: the feed pusher angle is changed from 90 degrees to 30 degrees to collect feed and then unload it, where 90 degrees represents collecting feed and 30 degrees represents unloading feed; operation description: first collect the scattered feed, and then push the feed to the foraging cows. The fourth feeding strategy includes: feeding plate unfolding angle: 90 degrees; operation description: clearing away the remaining feed; The fifth feeding strategy includes: feeding plate unfolding angle: 90 degrees; operation description: clear debris in the feed trough and send an alarm for insufficient feed to the system.

2. The method according to claim 1, characterized in that, The improved YOLOv10n model's neck network includes at least a first Zoom_cat stitching module, a second Zoom_cat stitching module, a dynamic scaling module, and an Add module; The first Zoom_cat stitching module is connected to the backbone network through a convolutional layer; The second Zoom_cat splicing module is connected to the backbone network through a convolutional layer, and is also connected to the first Zoom_cat splicing module through a convolutional layer. The dynamic scaling module is connected to the backbone network; The Add module is connected to the dynamic scaling module, and is connected to the second Zoom_cat splicing module through a convolutional layer.

3. The method according to claim 1, characterized in that, The training process of the image recognition model includes: Obtain an image dataset of the cow feeding area; The feed density in the image dataset is classified into several different feed density categories, which are then labeled. The cow foraging behaviors in the image dataset were classified to obtain multiple different cow foraging behaviors, and then labeled respectively. An improved YOLOv10n model is constructed based on the backbone network and the neck network, which is at least composed of the Zoom_cat splicing module, the dynamic scaling module, and the Add module. The labeled image dataset is input into the improved YOLOv10n model, and the improved YOLOv10n model is trained until the first recognition result of the improved YOLOv10n model reaches the first predetermined recognition result and the second recognition result reaches the second predetermined recognition result. The improved YOLOv10n model is then used as the image recognition model.

4. The method according to claim 3, characterized in that, The labeled image dataset is input into the improved YOLOv10n model, and the training of the improved YOLOv10n model includes: The labeled image dataset is input into the backbone network; The backbone network performs layer-by-layer downsampling on the input image dataset to obtain multiple first feature maps of different layers and sizes. The dynamic scaling module of the neck network dynamically enhances part of the first feature map and outputs multiple second feature maps. The first Zoom_cat stitching module of the neck network performs the first adaptive max pooling and average pooling adjustment on a portion of the first feature map to obtain the third feature map; The second Zoom_cat splicing module of the neck network performs a second adaptive max pooling and average pooling adjustment on some of the first and third feature maps to obtain the fourth feature map. The first target feature map is obtained by adding multiple second and fourth feature maps through the Add module of the neck network. The second target feature map is obtained by convolution and concatenation of multiple second, third, and fourth feature maps. The feature map obtained by downsampling the second target feature map is concatenated with the feature map output by the backbone network, and then a depthwise separable convolution operation is performed with one of the second feature maps to obtain the third target feature map.

5. The method according to claim 4, characterized in that, Multiple first feature maps of different layers and different sizes include at least a first sub-feature map, a second sub-feature map, and a third sub-feature map; the size of the first sub-feature map is larger than the size of the second sub-feature map, and the size of the second sub-feature map is larger than the size of the third sub-feature map. Furthermore, the first Zoom_cat concatenation module of the neck network performs the first adaptive max pooling and average pooling adjustment on a portion of the first feature map to obtain the third feature map, including: The size of the first sub-feature map is adjusted by the first Zoom_cat stitching module of the neck network so that the size of the first sub-feature map is the same as the size of the second sub-feature map. The size of the third sub-feature map is adjusted using the nearest neighbor interpolation method so that the size of the third sub-feature map is the same as that of the second sub-feature map; The adjusted first, second, and third sub-feature maps are spliced ​​together by channel dimension to obtain the third feature map.

6. The method according to claim 5, characterized in that, The dynamic scaling module of the neck network dynamically enhances a portion of the first feature map, outputting multiple second feature maps, including: The dynamic scaling module of the neck network is used to enhance the first sub-feature map by utilizing the first, second, and third sub-feature maps to obtain the enhancement result. The enhanced result, the original first sub-feature map, the original second sub-feature map, and the original third sub-feature map are respectively used as the second feature map and output.

7. The method according to claim 6, characterized in that, The dynamic scaling module of the neck network enhances the first sub-feature map using the first, second, and third sub-feature maps, resulting in the following enhancement: The dynamic scaling module of the neck network adjusts the first, second, and third sub-feature maps to the same spatial size. The size weights are dynamically generated for the adjusted first sub-feature map, second sub-feature map, and third sub-feature map using a preset network. By using dynamically generated size weights, the adjusted first sub-feature map, second sub-feature map, and third sub-feature map are fused to generate fused information, and the enhancement result includes the fused information.

8. A control system for intelligent feeding in a ranch, characterized in that, include: The acquisition module is configured to use a camera component located above the cow feeding area to acquire images of the cow feeding area and obtain image data; The recognition module is configured to input the image data into a trained image recognition model, and to identify feed density and cow foraging behavior in the image data through the image recognition model, and output a first recognition result corresponding to feed density and a second recognition result corresponding to cow foraging behavior; the image recognition model includes an improved YOLOv10n model; The generation module is configured to generate a corresponding material pushing strategy based on the first identification result and the second identification result; The control module is configured to control the pushing robot to perform pushing operations according to the pushing strategy; The generation module is further configured as follows: If the first identification result is sufficient feed, then the first feeding strategy is generated; If the first identification result is medium feed, then a second feeding strategy is generated; If the first identification result is a small amount of feed and the second identification result is that dairy cows are foraging, then a third feed pushing strategy is generated. If the first identification result is a small amount of feed and the second identification result is no dairy cows foraging, then a fourth feed pushing strategy is generated. If the first identification result is no feed, then the fifth feeding strategy is generated; The first feeding strategy includes: feeding plate unfolding angle: 30 degrees; feeding robot walking direction: along the parallel direction of the feed trough; operation description: push excess feed from the outside of the feed trough back into the feed trough. The second feeding strategy includes: the feeding plate unfolding angle: 45 degrees; the feeding robot walking direction: along the parallel direction of the feed trough; operation description: push the feed outside the feed trough back into the feed trough, and adjust the feed distribution to make the feed in the feed trough evenly distributed. The third feeding strategy includes: the feed pusher angle is changed from 90 degrees to 30 degrees to collect feed and then unload it, where 90 degrees represents collecting feed and 30 degrees represents unloading feed; operation description: first collect the scattered feed, and then push the feed to the foraging cows. The fourth feeding strategy includes: feeding plate unfolding angle: 90 degrees; operation description: clearing away the remaining feed; The fifth feeding strategy includes: feeding plate unfolding angle: 90 degrees; operation description: clear debris in the feed trough and send an alarm for insufficient feed to the system.

Citation Information

Patent Citations

  • Auxiliary feeding robot for dairy cow and auxiliary feeding method

    CN112975962A

  • Method and system for predicting feed intake of perinatal cows in pasture breeding scene

    CN119168425A

  • Improved Yolov10-oriented target detection fusion method

    CN119723272A