The planting density of rice in the early seedling stage is closely related to yield estimation and field management. With the development of computer vision technology, the use of drone image recognition technology to automatically count rice is gradually replacing traditional manual sampling and counting methods. However, images taken by drones have problems such as irregular rice seedling shapes, overlapping rice seedlings, scale changes, and background occlusion, which lead to deviations in feature learning, which in turn affects the accurate estimation of rice density and the efficiency of computational prediction. Significant relative size differences among rice seedlings will also become another important factor affecting counting accuracy. In response to these problems, a rice seedling counting algorithm (rice counting network, RCN) was proposed based on an improved multi-column convolutional neural network. Based on the size and shape of the target, an improved Gaussian density kernel was proposed, which combined the similarity information of the rice image and the RGB image of the center point to make the density kernel more suitable for the texture of the rice seedlings. The number of rice ears of different sizes and shapes was quantified, and the scales were divided finely to calculate the receptive field sizes of different features that needed to be fused. The fusion expansion convolution module was used to increase the receptive field size and enhance the model's multi-scale capabilities. On this basis, appropriate attention mechanisms were added to the pathways at different scales to enhance attention to reasonable information about rice seedlings at different scales. It showed obvious effectiveness in scenes with complex backgrounds, multi-scales and dense occlusions. On the RSD2023 data set captured at a height of 5 m, RCN's mean absolute error MAE=3.10 holes/m² and mean square error MSE=4.03 holes²/m⁴, which were better than that of the mainstream algorithms such as MCNN, CSRNet, SCAR and TasselNetV2.

Deep learning methods are widely applied in agricultural vision tasks. Determining harvesting parameters is a primary task for the mechanized harvesting of Camellia oleifera fruits, and automated fruit detection and object counting based on deep learning are crucial for accurately setting these parameters. However, factors such as occlusion and illumination variations in complex orchard environments directly affect the efficiency and accuracy of detection methods. An efficient, lightweight model named YOLO-LGC was proposed based on YOLO 11n, aiming to enhance target detection accuracy and spatial localization capability in complex orchard environments. Firstly, the LNN ghost dynamic convolution (LGC) module was introduced into the backbone network to optimize feature extraction through adaptive kernel selection, thereby improving detection accuracy in complex environments. Secondly, the programmable gradient information (PGI) mechanism was adopted to dynamically adjust gradient propagation, addressing feature conflicts caused by occlusion and illumination changes. Finally, the Dynamic UpSample module was integrated, guided by PGI, to enhance feature retention and improve spatial localization accuracy. Combined axis-aligned bounding boxes (AABB), an inertial measurement unit (IMU), and Kalman filtering to visualize the spatial distribution of Camellia oleifera fruits at heights of 0~1.5 m, 1.5~1.8 m, and above 1.8 m. By analyzing the spatial relationships of the fruits, more accurate positional information was provided. Experimental results showed that compared with YOLO 11n, YOLO-LGC improved detection precision from 82.0% to 83.0%, recall from 76.4% to 78.0%, mAP@0.5 from 85.1% to 87.6%, and FPS from 229.8 f/s to 299.7 f/s, with computational complexity slightly increased to 6.7×10⁹. While maintaining the advantage of high inference speed, YOLO-LGC enhances detection performance, outperforming existing mainstream detection methods. The spatial distribution error rate was less than 5%, meeting practical detection requirements. The findings can provide a basis for matching mechanized harvesting parameters for Camellia oleifera fruits.

Aiming at the problems such as occlusion and large deformation of cow images in the cowshed under a wide field of view, which cause difficulties in individual identification, a method for individual identification of cows in the cowshed under a wide field of view based on the DAM-ResNet model was proposed. The Mask R-CNN model was used to segment four types of dairy cow parts: the back, the left side of the trunk, the right side of the trunk and the buttocks. Based on the ResNet34 residual network, the second-generation deformable convolution was introduced to enhance the extraction of deformed image features of cow patterns in the segmentation results. Integrating the AFF attention feature fusion module into the residual structure to achieve accurate recognition of images of small target cows at a distance. The fine-grained classification loss function, interchannel loss (MC-Loss), was adopted to improve the recognition accuracy of the model for cows with similar patterns. The multi-part dataset of dairy cows was constructed by using the segmented images of dairy cows, and the DAM-ResNet model was trained. Individual cow recognition tests were conducted on a dataset of 12 864 images of 57 dairy cows. The results showed that the recognition accuracy rates of the back, left trunk, right trunk and buttocks of dairy cows were 96.13%, 96.56%, 96.94% and 93.14%, respectively, which were 2.76, 2.88, 2.92 and 4.25 percentage points higher than those of the original ResNet model. The recognition accuracy rates of the method proposed within the ranges of 10~20 m, 20~30 m and 30~40 m were 97.54%, 90.72% and 82.17%, respectively. The research results can provide technical support for intelligent dairy cow breeding.

With the intelligent development of animal husbandry, computer-vision-based individual cattle recognition is playing an increasingly important role in modern cattle farming. However, in practical applications, the rapid changes in the angles of cattle faces, frequent fluctuations in the number of cattle, and limitations in model recognition speed severely affect the accuracy and efficiency of recognition. To address these challenges, an individual recognition method for Simmental cattle was proposed based on improved YOLO v8 keypoint detection. The method consisted of two stages: in the cattle face detection stage, the improved YOLO v8 model automatically detected keypoints in cattle face images and performed correction alignment and background cropping to make the angles of the cattle face images more consistent. In the cattle face recognition stage, depthwise separable convolution was firstly introduced to reduce the number of parameters, and then residual connections and attention mechanisms were utilized to enhance the model's feature extraction capabilities, thereby improving recognition accuracy and speed. To validate the effectiveness of the method, a dataset of cattle face images containing 159 Simmental cattle was constructed. Experimental results showed that the proposed model had significant improvements in the number of parameters and inference speed, with processing time for a single image reduced to within 39 ms and a recognition accuracy of 95.8%, an average improvement of 4.8 percentage points compared with that of other models. Additionally, the proposed method supported real-time updates to the cattle database, maintaining high accuracy even after adding new cattle face images.

Given the current research on beef cattle behavior, which mainly focused on basic behavior recognition and lacked local perception of complex behaviors, the identification of fine behaviors under basic behaviors of beef cattle was researched. A multi-behavior detection and statistical method based on YOLO v8 for beef cattle was proposed. Cattle behavior images were collected by cameras to build a comprehensive dataset that included fundamental behaviors such as standing, lying down, feeding, and drinking, as well as fine behaviors like licking, walking, searching, and tail flicking. YOLO v8n-P2 was selected as the basic model to enhance the ability of the model to detect calves; the feature extraction structure of C2F-PPA was designed; the YOLO v8 detection head was improved by interactive strategy, and the TDADH was constructed. MPDIoU loss function was used to address limitations associated with CIoU. Subsequently, statistical analysis of cattle behaviors based on detection results throughout the day was conducted; these results were visually displayed for various occlusion environments. The experimental results showed that the PTPM-YOLO v8n model achieved precision rate of 91.0%, a recall rate of 87.9%, and a mAP@0.5 score of 94.3% in recognizing all eight behaviors tested. Compared with the original model YOLO v8n, the mAP@0.5 of PTPM-YOLO v8n was increased by 3.0 percentage points, and the parameter number was decreased by 21.9%. which identified all behaviors and basic behaviors, the mAP@0.5 was increased by 3.0 and 2.4 percentage points, respectively. The method presented can accurately identify fine behaviors of beef cattle under farming conditions, providing a reference for multi-behavior monitoring of beef cattle.

In intelligent cattle farming, the application of facial recognition technology for cattle is crucial. However, due to the complex and variable farming environment and the non-planar characteristics of cattle faces, facial recognition technology often faces challenges such as insufficient recognition accuracy and poor robustness in practical applications. To address these issues, a FaceNet based on multi-scale fusion network (FMF) was proposed, which identified cattle through facial features. Firstly, the MSRCR method was used for color restoration preprocessing of the input cattle face images to reduce the impact of lighting on the FMF algorithm. Subsequently, a more lightweight MobileNetV3 was introduced into the main feature extraction network to reduce the model parameters and computation while ensuring high feature extraction capability. Finally, a composite dual-branch adaptive attention mechanism (M_CBAM) was proposed for multi-scale feature fusion. M_CBAM adjusted the weighting coefficients based on important features of the feature maps, adaptively weighting serial CBAM and parallel CBAM weights, then performed multi-scale feature fusion of local fine features and global features of cattle faces, improving facial recognition accuracy. To explore the effectiveness and real-time performance of the proposed algorithm, ablation experiments were conducted on a self-made cattle face dataset, and the results were compared with current mainstream recognition algorithms. Finally, the algorithm was deployed on the Jetson AGX Xavier embedded platform for application testing. The test results showed that the proposed algorithm achieved an accuracy of 93.86% and an FPS of 30.02 f/s on an open test set collected from other dairy farms. Under the condition of fast model inference speed, the recognition accuracy was significantly better than that of the original network and comparison networks.

The accurate measurement of granular fertilizer mass flow rate presents a significant challenge for achieving precision fertilization. Due to the complex dynamics of granular flows, traditional measurement methods often struggle to handle the resulting signals, which are typically characterized by significant noise and non-stationarity. To address this issue, a novel pre-processing algorithm for granular fertilizer flow signals acquired by a microwave sensor was proposed. Initially, a model of the fertilizer discharge pipeline was constructed and simulation experiments were conducted. These simulations determined that under conditions of a 400 mm pipeline length and 30 mm diameter, the impact of particle collisions within the pipeline on velocity was minimized. Based on these simulation results, a signal acquisition platform for granular fertilizer mass flow was established. Through analysis of the acquired flow signals, a hybrid denoising algorithm combining variational mode decomposition (VMD) and discrete wavelet transform (DWT) was innovatively proposed. Experimental results demonstrated that after processing with this algorithm, the signal-to-noise ratio of the signal was increased by 7.375 7 dB, and the root mean square error was decreased by 57%. The algorithm effectively separated the noise component from the signal, thereby significantly improving the measurement accuracy of the granular fertilizer mass flow rate. Finally, real-world vehicle tests were conducted. The processed flow signals exhibited a maximum relative error of 13.84% and a minimum relative error of 1.36%. The results confirmed that the experimental platform developed and the proposed denoising algorithm provided reliable technical support for precision fertilization.

In order to improve the accuracy of unmanned agricultural machine's perception of obstacles in the farm environment, to solve the problem that visual detection is easily affected by light and millimeter-wave radar detection is easily affected by vehicle bumps, etc., as well as the problem that the visual target detection algorithm has a large number of parameters, a large amount of computation, and a large volume of the model under the complex field, this paper proposes an obstacle-avoidance target detection method for unmanned agricultural machines with the fusion of visual and millimeter-wave radar information. Part of the radar target data from millimeter-wave radar is first filtered, and a target tracking algorithm based on adaptive extended Kalman filtering is proposed. Then a farm environment obstacle dataset is produced and a target detection network based on improved YOLO v5s is constructed. Subsequently, the mapping of radar points into the image pixel coordinate system is realized by time-stamp alignment and coordinate transformation with direct linear calibration method. Finally, the fusion model of obstacle detection information from millimeter-wave radar and visual sensors is constructed through the decision-level fusion method and target matching strategy, and the experimental results show that the mean average accuracy of the improved YOLO v5s is 97.0%, which is similar to that of the original model, but the number of parameters, the amount of computation, and the size of the model are only 40.2%, 39.2%, and 38.2% of the original YOLO v5s model, respectively. Compared with YOLO v4-Tiny, YOLO v7-Tiny, YOLO v4 and YOLO v7 models, it can better balance the detection accuracy and speed. The results of multi-scene tests show that the fusion method proposed in this paper improves the recognition accuracy by 2.67 percentage points and 15.07 percentage points compared with radar and camera in daytime tests, and the fusion detection method can effectively make up for the failure of the camera in nighttime tests, and has better robustness and accuracy than the single-sensor algorithm, and the fusion obstacle avoidance system effectively realizes the parking avoidance of the unmanned agricultural machine.

Aiming to address the core challenges faced by agricultural robots in autonomous navigation and path planning under complex farmland environments—particularly in irregular plots and soft, sticky paddy fields where common issues included low coverage rate, excessive turning redundancy, and poor path continuity, a full-coverage path planning algorithm tailored for convex polygonal farmlands and compatible with dual-helix driven agricultural robots was proposed. The algorithm integrated both the geometric characteristics of the farmland and the kinematic constraints of the robot. The working area was initially dividedintoacentralparallel-workingzoneandaperipheralcontour-followingzone, corresponding respectively to strip-based traversal paths and boundary-parallel strip traversal paths. The complete path was further segmented into working and non-working paths, including center-region connection paths, inter-region transitions, contour-parallel region transitions, and entry/exit paths. The methodology for generating each type of path was detailed. An optimal working direction was determined via the “long-side-first” principle for the central region, and the Warnsdorff rule was applied to optimize the sequence of strip traversal. Dubins and Reeds-Shepp path models were introduced to enhance continuity and feasibility during turns. Considering metrics such as path length, computation time, and coverage rate, the algorithm was implemented on a dual-helix driven robot and tested in simulations and field experiments across four typical farmland types: rectangular, trapezoidal, irregular quadrilateral, and irregular polygonal plots. Experimental results demonstrated that the proposed algorithm achieved a coverage rate between 96.69% and 97.80%, with a path overlap rate controlled within 1.77%~2.30%. The generated paths were continuous and boundary-safe, indicating strong execution stability and environmental adaptability. The research result can provide a robust and adaptable path planning solution for agricultural robots operating in complex environments, with promising engineering feasibility and application potential.

Traditional visual navigation methods struggle to handle interference from varying lighting conditions and weed occlusion in complex pear orchard environments. An improved navigation line extraction method was proposed based on YOLO v8 to address this challenge. The process integrated a multi-scale large kernel attention (MLKA) module into the YOLO v8 model to enhance the perception of trunk features. A multi-frame feature point fusion mechanism was designed. By recording and utilizing feature points detected in five consecutive frames, this mechanism effectively compensated for the problem of insufficient feature points in single-frame detection. Additionally, the random sample consensus (RANSAC) algorithm was introduced to denoise the feature points of the left and right tree rows, respectively, and the least squares method was used for the two side tree rows' line fitting. The navigation line was ultimately generated by calculating the angle bisector of the fitted lines from both sides of the tree rows. Experimental results showed that the improved model achieved a precision of 89.8%, a recall of 79.9%, and a mean average precision (mAP50-95) of 55.1% in tree trunk detection tasks under various lighting conditions and weed occlusion scenarios. The navigation line generated by combining multi-frame feature point fusion with RANSAC denoising exhibited an average angular deviation of only 1.17° from the manually annotated reference navigation line, an average position deviation of 20.40 pixels, and an average root mean square deviation of 0.27. The research result can provide a low-cost and highly adaptable technical solution for visual navigation in pear orchard environments.