Global path planning is crucial for the autonomous navigation of orchard vehicles. Yet, how to leverage global structural information to manage task sequencing and generate coverage paths across the entire operation area remains underexplored. This study proposes a DBSCAN-OBB-based method to extract tree-row endpoints and automatically generate key waypoints constrained by vehicle parameters. The coverage operation is formulated as a kinematically constrained shortest-path problem, and the optimal waypoint sequence is obtained using the CGB algorithm. A global path planner, MultiPointPlanner, constructs and concatenates inter-row and headland segments, yielding smooth and feasible global paths. Experimental results show that the proposed method achieved a mean waypoint error of <= 0.19 m. For orchards of different scales, the CGB algorithm generated optimal waypoint traversal sequences, achieving up to 39.7% non-productive travel reduction compared to the row-by-row rule, while ensuring equivalent optimality within a computation time of <= 0.1 s. In path-tracking experiments, the planned global paths exhibit low curvature fluctuation and consistent heading. In headland regions, the proposed planner keeps trajectories away from high-risk boundaries, allowing the vehicle to follow smoothly under continuous curvature constraints. The row-switching time is reduced by 61.1%, and the minimum obstacle clearance is increased by 64.0%. Across the entire operation area, the number of reversals and manual takeovers was substantially reduced, confirming the effectiveness of the proposed approach in achieving geometrically smooth, kinematically feasible, and risk-avoiding coverage paths.

Cleaning-loss strongly affects rice combine harvesting efficiency, yet piezoelectric monitoring can become unreliable when the machine transitions between wet, dry, and combined paddy zones. This study developed a field-condition-aware lightweight 1D CNN_GRU virtual sensor to estimate voltage-based cleaning-loss intensity using machine-dynamics measurements. A multi-sensor platform was deployed on a ZOOMLION rice combine harvester, integrating dual IMUs (sieve and chassis), fan/rotor proximity RPM sensors, field-measured agronomic variables, and RTK-GNSS, while piezoelectric thin-film array and ceramic sensors provided training targets. Raw piezoelectric signals were sampled at 1000 Hz and down sampled to a synchronized 1 Hz logging rate for multi-sensor fusion. Field data were collected across wet, dry, and combined field conditions (15,000 time-aligned samples retained for each condition). Indoor grain-dropping calibration test confirmed the quantitative relationship between known grain impact mass and piezoelectric voltage response. The proposed 1D CNN_GRU model (∼73 k parameters, ≈ 0.28 MB) was benchmarked against machine-learning (ML) baselines (Ridge, SVR, Random Forest, XGBoost) and deep-learning (DL) baselines (1D_CNN, GRU, LSTM, CNN_LSTM, CNN_LSTM_GRU). The proposed model outperformed traditional ML and DL baselines, achieving (Coefficient of Determination (R2) / Root Mean Squared Error (RMSE) values of 0.87/0.238 V and 0.90/0.211 V; 0.96/0.013 V and 0.97/0.012 V; 0.98/0.126 V and 0.98/0.105 V) under wet, dry and combined field conditions for the array and ceramic sensors, respectively. Under wet conditions, the virtual sensor outperformed the best ML baseline XGBoost (ΔR2 = +0.25 for array and ΔR2 = +0.23 for ceramic). Sensor ablation and SHAP analysis identified field condition dependent dominant predictors, while MC Dropout provided uncertainty aware prediction intervals. These findings demonstrate that the proposed virtual sensor provides an accurate and redundant cleaning-loss intensity estimation channel that remains operational when direct piezo measurements become unreliable, supporting embedded-oriented monitoring for smart rice harvesting.

Accurate terrain classification is essential for the effective and efficient operation of autonomous robots and off-road vehicles. This study proposes a novel multi-sensor fusion technique for off-road terrain classification using a robotic platform that matches the practical features of an electric tractor. The method employs proprioceptive signals from the vehicle body and all four wheels, including wheel torque, wheel speed, wheel slippage, and vibration. A high-speed counter was programmed using the STEP 7 Microwin environment and uploaded onto an S7-200 PLC to measure wheel torque and speed. An inertial measurement unit and a three-axis digital accelerometer were used to record vibrations from the vehicle body and wheel bracket, respectively. MATLAB Simulink was employed to synchronise sensors data. Signal pre-processing, segmentation, Savitzky-Golay filtering, feature-level fusion, and the Random Forest algorithm were incorporated into the developed terrain classification model. The proposed Random Forest-based model, validated using K-fold cross-validation, achieved up to 90% test accuracy. Performance on unseen labelled data demonstrated consistent classification accuracy between 80% and 90%, indicating strong generalisation across terrain types. Real-time testing with unseen and unlabelled data demonstrated the model’s robustness and stability, enabling reliable terrain prediction with minimal latency, suitable for real-world off-road autonomous vehicle applications.

Integrated puddling-leveling operation is a critical step in paddy field preparation, typically conducted between plowing and rice transplanting. However, the accuracy of elevation measurements in existing automatic leveling technologies is often constrained by limited operating ranges or susceptibility to electromagnetic interference, resulting in inconsistent leveling performance. Because the water surface naturally reflects terrain undulations in paddy fields, this study proposes a semantic segmentation-based approach to detect exposed soil regions for guiding a floating-type puddling and leveling implement. To this end, a lightweight semantic segmentation model, PL_DeepLabV3+_0.8, was developed specifically for integrated puddling-leveling operation. The model combines a MobileNetV2_S backbone, a Low-Level Feature Fusion Module (LFM), and structured pruning. These components collectively enable the rapid and accurate detection of exposed soil in paddy fields under computationally constrained conditions. The PL_DeepLabV3+_0.8 model was successfully deployed in the control system of a floating-type implement, and its effectiveness was validated through field tests conducted at different operating speeds and modes. On a paddy field image dataset, PL_DeepLabV3+_0.8 achieved a mean Pixel Accuracy (mPA) of 92.23 +/- 0.22%, a mean Intersection over Union (mIoU) of 84.18 +/- 0.31%, and an inference speed of 7.73 frames per second (FPS), outperforming the original DeepLabV3 + model, which achieved 91.90%, 83.81%, and 0.88 FPS, respectively. In field tests at operating speeds of 1.1 m/s and 1.5 m/s, the surface flatness (standard deviation of elevation) in two paddy fields was improved from 3.61 cm and 4.07 cm to 2.11 cm and 2.42 cm, respectively. These results indicate that the deployed model not only satisfies the flatness requirement for rice transplanting (< 3 cm) but also delivers a productivity increase of 0.28 ha/h compared with conventional manual operation. Overall, this study provides a useful reference for the development of intelligent puddling and leveling technologies in paddy field preparation.

Reliable obstacle perception and motion prediction are essential for the safe autonomous operation of agricultural robots in nighttime farmland environments. To address the degradation of visible-light perception under low illumination, this study proposes a framework integrating thermal imaging and LiDAR for obstacle detection, 3D localization, and trajectory prediction. First, YOLOv10n detects three typical categories of potentially dynamic agricultural obstacles in thermal images: persons, tractors, and weeders. On an independently collected test set, YOLOv10n achieves a precision of 87.2%, a recall of 80.4%, an mAP@50 of 86.2%, an mAP@50–95 of 57.5%, and an inference speed of 152 FPS. Second, thermal-image bounding boxes are associated with LiDAR point clouds through RANSAC-based near-ground point suppression, image–point cloud projection, DBSCAN clustering, and vision-guided optimal cluster matching. A depth-weighted median method estimates obstacle positions, achieving a maximum localization RMSE of 0.1706 m over the evaluated range. Finally, an Improved Agricultural Interacting Multiple Model (IA-IMM) framework is developed for trajectory prediction using obstacle-category-based motion model configurations, kinematic consistency constraints, and farmland traversability priors. Experiments on 30 real farmland obstacle trajectories show that, for a six-step prediction horizon, IA-IMM reduces average displacement error (ADE) by 10.49% and RMSE by 14.14%, while increasing Success@0.2m by 6.4 percentage points compared with the standard IMM. These results demonstrate the effectiveness of the proposed framework for nighttime obstacle perception, localization, and trajectory prediction in farmland environments.

Currently, unmanned rice harvesters operating at a fixed speed suffer from limited adaptability and autonomy. We proposed an operating speed prediction and control method for the harvester based on multi-source information and data amplification. We established an operation speed control platform for the harvester, built an "end-to-end" prediction model from the harvester's multi-source information perception to the operation speed. Aiming at the objective fact that it is difficult to obtain agricultural machinery data samples, we proposed a GANbased multivariate agricultural sequence data amplification method. Finally, we optimized the operating speed of the harvester, developed a speed control strategy, and conducted field tests. The results indicate that after GAN data amplification, the speed prediction model's coefficient of determination (R2) reached 0.8696, and the mean squared error (MSE) reached 0.0466. Compared to the original dataset, R2 improved by 14.74 %, and MSE decreased by 23.48 %. In the field tests of speed control, the standard deviations of the cruising speed in two different paddy fields were 0.26 km/h and 0.18 km/h. Furthermore, the speed variation exhibited strongly correlated with changes in perceived information, and the drum speed remained stable without obvious "jerking" or clogging. These findings suggest that the speed prediction and control system operates reliably and can serve as a reference and provide technical support for research on autonomous operation of unmanned rice harvesters.

Varieties show their unique characteristics in morphology, growth, and fruits. Tomato maturity is related to multiple dimensional characteristics including color, texture, smell, etc. An effective classification method of tomato variety and maturity is crucial for evaluating its growth and yield. However, due to the complex growth environment, some problems such as leaf occlusion and fruit shaded by each other make it difficult to accurately and efficiently identify them. To solve these problems, this study innovatively proposes a simultaneous detection model on tomato variety and maturity based on improved YOLOv8n, with the combination of frequency-adaptive dilated convolution (FADC) feature extraction module and the high-level screening-feature path aggregation network (HSPAN) with the aim of local and global feature fusion by the channel attention module and feature selection fusion mechanism. In addition, we use the Powerful-IoU (PIoU) loss function to replace the original Complete IoU (CIoU) to enhance the accuracy of bounding boxes. We also introduce a dynamic detection head as the final output of the model, which can adaptively adjust the focus of feature extraction according to the color and size of tomato fruits, thereby improving the recognition accuracy. Experimental results show that our model with better global perception capability achieves the highest detection accuracy and lower computation complexity among the comparative models.

Effective obstacle detection is vital for autonomous agricultural robots in complex environments. Due to tree occlusion and varying lighting conditions, current systems need help with misidentifying obstacles, particularly in orchards. This study presents a novel methodology for real-time obstacle detection and distance estimation, improving the YOLOv8n architecture with the Convolutional Block Attention Module (CBAM) to improve feature representation. Additionally, the Non-Maximum Suppression (NMS) technique is enhanced with Soft DIou-NMS to minimize redundant detections for overlapping objects. A comprehensive dataset of common orchard obstacles was utilized for evaluation, measuring performance through Precision (P), Recall (R), F1-score, and confusion matrices. Experimental results showed that the improved YOLOv8n model outperforms baseline models, including YOLOv5s and various YOLOv8 variants (n, s, m, l, x), achieving a 92.7 % mean Average Precision (mAP) at IoU 0.5 and an F1-score of 87 % at a confidence threshold of 42.1 %. Inference time was reduced to 3.3 ms (ms), and the model size to 20.1 MB. The model was evaluated under various lighting conditions, achieving mAP-50 scores of 96.5 % in daylight and 91.4 % in low light, with minimal performance drops. Testing across distance ranges (close: 1-3 m, medium: 3-7 m, long: 7-12 m) showed strong performance in close and medium ranges. Real-time validation indicated a detection time of 20.1 ms per frame, making it suitable for agricultural vehicles. These findings provide a reference standard for obstacle detection and distance estimation in orchard environments, contributing to the safety and autonomy of agricultural vehicles in complex settings.

Autonomous navigation in agricultural environments requires precise obstacle classification to ensure collision-free movement. This study proposes a convolutional neural network (CNN)-based model designed to enhance obstacle classification for agricultural robots, particularly in orchards. Building upon a previously developed YOLOv8n-based real-time detection system, the model incorporates Ghost Modules and Squeeze-and-Excitation (SE) blocks to enhance feature extraction while maintaining computational efficiency. Obstacles are categorized as “Real”—those that physically impact navigation, such as tree trunks and persons—and “Fake”—those that do not, such as tall weeds and tree branches—allowing for precise navigation decisions. The model was trained on separate orchard and campus datasets and fine-tuned using Hyperband optimization and evaluated on an external test set to assess generalization to unseen obstacles. The model’s robustness was tested under varied lighting conditions, including low-light scenarios, to ensure real-world applicability. Computational efficiency was analyzed based on inference speed, memory consumption, and hardware requirements. Comparative analysis against state-of-the-art classification models (VGG16, ResNet50, MobileNetV3, DenseNet121, EfficientNetB0, and InceptionV3) confirmed the proposed model’s superior precision (p), recall (r), and F1-score, particularly in complex orchard scenarios. The model maintained strong generalization across diverse environmental conditions, including varying illumination and previously unseen obstacles. Furthermore, computational analysis revealed that the orchard-combined model achieved the highest inference speed at 2.31 FPS while maintaining a strong balance between accuracy and efficiency. When deployed in real-time, the model achieved 95.0% classification accuracy in orchards and 92.0% in campus environments. The real-time system demonstrated a false positive rate of 8.0% in the campus environment and 2.0% in the orchard, with a consistent false negative rate of 8.0% across both environments. These results validate the model’s effectiveness for real-time obstacle differentiation in agricultural settings. Its strong generalization, robustness to unseen obstacles, and computational efficiency make it well-suited for deployment in precision agriculture. Future work will focus on enhancing inference speed, improving performance under occlusion, and expanding dataset diversity to further strengthen real-world applicability.

The motion of a robot in the orchard environment is based on the identification of obstacles in the orchard environment. To achieve this, the robot should identify the obstacles and navigate away from such obstacles. This study proposes a method using LiDAR data to identify obstacles and process the location of all the obstacles. The method comprises of three key components: identification and segmentation of apple canopy features, identification of obstacles and trunks, mapping of the location of all obstacles and trunks. Firstly, the canopy features are identified based on the criteria; height, density and intensity. The identified canopy is bounded by voxels that delineate the boundaries of the canopy. Secondly, the obstacles and trunks are mapped, resulting in a guideline for the navigation of the orchard robot. The experimental results show that the canopy feature removal reduced processing time by 63.7 % compared to processing the full point cloud and was 42.3 % faster than Euclidean clustering method, significantly lowering computational demands. The mapping of trunks achieved an accuracy of 92.3 % by locating all the positions of the apple trunks. The experimental results yielded an average mapping error of 0.034m. The proposed algorithm in this study generates an orchard map of trunk features enabling precision agriculture applications such as yield prediction which is achieved through trunk canopy correlation and autonomous navigation.

The accurate localization of fruits and vegetables is crucial for robotic harvesting systems, depending not only on the localization accuracy within the vision device frame but also on the precision of the hand-eye pose. Despite considerable research focusing on localization accuracy within the vision device frame, limited attention has been directed toward improving the accuracy of the hand-eye pose. Furthermore, current methods for acquiring an accurate hand-eye pose are often inflexible and difficult to apply for the users. Therefore, we propose an accurate and flexible hand-eye pose acquisition method to enhance the performance of an apple fruit localization system. This method begins by attaching a datum block with three marked points to the camera, and the camera pose relative to the datum block is then determined using an iterative least-squares algorithm in the proposed calibration system. Once calibrated, the camera with the attached datum block can be mounted on a harvesting robot. In this manner, users can easily acquire the hand-eye pose by introducing the robotic arm of the harvesting robot along with a probe to contact the marked points on the datum block, making the process more flexible and accessible. To validate the effectiveness of the proposed hand-eye acquisition method, extensive experiments were conducted using a self-designed apple harvesting robot. Comparative experiments demonstrated that our method significantly improved the accuracy of the fruit localization system, reducing the mean positioning error by at least 33.3% compared to the selected state-of-the-art methods. Overall, the system achieved a mean localization accuracy of 14.0 mm under non-occluded conditions, with illuminance levels ranging from 8 to 20 kLux in a simulated orchard environment. Furthermore, field evaluation experiments in an apple orchard revealed that the localization success rate of the system exceeded 80.53%.

The fusion of the hyperspectral image (HSI) and light detection and ranging (LiDAR) data has demonstrated significant potential in the land cover classification task. Although deep learning has shown remarkable success in the joint classification of HSI and LiDAR data, the large amount of unlabeled multisource remote sensing data is not fully utilized. Additionally, effectively integrating HSI and LiDAR data remains a challenging task, and the semantic relationship of neighborhood regions needs to be further exploited. In this paper, we propose a cross-source transformer-based neighborhood contrastive learning model (CTNCLM), which acquires a more discriminative feature representation from unlabeled data through the pre-training stage. Considering the semantic correlation between neighboring image patches, CTMCLM achieves the joint classification of HSI and LiDAR data at a more precise level. A cross-patch contrastive learning (CPCL) module is proposed to calculate the similarity between original patches and neighborhood patches. Furthermore, a cross-source transformer (CST) with cross-source attention is proposed to fuse the multi-source data, which exploits the intermodal information interaction between the HSI and LiDAR data. Extensive experiments on three public datasets demonstrate the superior classification performance of the proposed method compared with several state-of-the-art methods.

Field crops are usually planted in rows, and accurate identification and extraction of crop row centerline is the key to realize autonomous navigation and safe operation of agricultural machinery. However, the diversity of crop species and morphology, as well as field noise such as weeds and light, often lead to poor crop detection in complex farming environments. In addition, the curvature of crop rows also poses a challenge to the safety of farm machinery during travel. In this study, a combined multi-crop row centerline extraction algorithm is proposed based on improved YOLOv8 (You Only Look Once-v8) model, threshold DBSCAN (Density-Based Spatial Clustering of Applications with Noise) clustering, least squares method, and B-spline curves. For the detection of multiple crops, a DCGA-YOLOv8 model is developed by introducing deformable convolution and global attention mechanism (GAM) on the original YOLOv8 model. The introduction of deformable convolution can obtain more fine-grained spatial information and adapt to crops of different sizes and shapes, while the combination of GAM can pay more attention to the important feature areas of crops. The experimental results shown that the F1-score and mAP value of the DCGA-YOLOv8 model for Cabbage, Kohlrabi, and Rice are 96.4%, 97.1%, 95.9% and 98.9%, 99.2%, 99.1%, respectively, which has good generalization and robustness. A threshold-DBSCAN algorithm was proposed to implement clustering for each row of crops. The correct clustering rate for Cabbage, Kohlrabi and Rice reaches 98.9%, 97.9%, and 100%, respectively. And LSM and cubic B-spline curve methods were applied to fit straight and curved crop rows, respectively. In addition, this study constructed a risk optimization function for the wheel model to further improve the safety of agricultural machines operating between crop rows. This indicates that the proposed method can effectively realize the accurate recognition and extraction of navigation lines of different crops in complex farmland environment, and improve the safety and stability of visual navigation and field operation of agricultural machines.

The robot gripper, as an interface for physical-information interaction between agricultural robots and the operating environment, has been widely used in agricultural production. The potential slipping risk during the grasping process is an important factor affecting safe gripping. Therefore, detecting the initial slipping during the gripping process and optimizing the force applied during gripping are key technologies for preventing slippage and achieving safe gripping. This paper proposes a slip detection method based on tactile sensing and time–frequency analysis, it can attain slip signal detection and grip force optimization during fragile fruits gripping. We have constructed a compliant robotic hand grasping system and developed a tactile information acquisition platform based on contact force sensors and bending sensors. This platform is capable of real-time monitoring and recording tactile sensing data during grasping experiments. To deeply analyze the tactile temporal information during the grasping process, we employed the Short-Time Fourier Transform (STFT) and Discrete Wavelet Transform (DWT) algorithms, which precisely captured the variations in the contact force signal in the time–frequency domain during the sliding event. With the application of machine learning techniques, we not only achieved effective slip detection but also optimized the control of the grasping force, providing valuable references and guidance for non-destructive grasping technology in the field of agricultural robotics.

Rice fertilization management plays an important role in rice yield and quality; however, making automatic fertilization plans according to the rice life cycle is difficult. This study proposes a rice fertilization recommendation model using a knowledge graph and a case-based reasoning method. To build the recommendation model, an initial fertilization scheme was first obtained by retrieving information from a knowledge graph. It was further enriched through operating vectors of all entities and relations in the knowledge graph obtained using the PairRE model. Next, k historical cases similar to the target event were retrieved using the fertilization scheme, attributes of the target event, and entity vectors of the knowledge graph. Finally, the fertilization rates for the target were predicted according to the k cases. The results showed that by employing the proposed model, the error rate of the fertilization scheme was approximately 19.5% and the prediction accuracies for N, P2O5, and K2O were 95.69%, 79.04%, and 81.67%, respectively. The model not only improves the accuracy of recommendations but also has strong interpretability and high practical value, and can be easily replicated and applied in the design of agricultural recommendation systems.

Highlights Response surface methodology is suitable for DEM input parameter optimization. Soil reaction forces reduced at velocity ratios greater than one (1.2-3.9). Vibration reduced soil reaction forces at the target depth of 350 mm by 70%. In general, soil reaction forces increase with speed but decrease with frequency. Abrasive wear predominantly occurred at the tool’s cutting section. Abstract. The discrete element method (DEM) and response surface methodology (RSM) were used to determine the input parameters and combination of operational factors required for optimizing the Jerusalem artichoke (Helianthus tuberosus L.) harvesting tool in cohesive soil. The DEM soil model consisted of particles with different radii in three shapes calibrated using angle of repose and cone penetration data. Compared with data from a soil bin subsoiler evaluation, the DEM model showed acceptable relative errors for draught force (6.7%), vertical force (4.5%), and furrow width (9.3%). The effects of operational factors, including forward speed, vibration frequency, and amplitude, on response variables such as draught and vertical forces, drawbar power, and abrasive wear were analyzed for three harvesting shovels (S-shape, step-shape, and fork-shape). The ratio of vibratory speed to forward speed (velocity ratio, Vr) was used to analyze the combined effect of the factors. The operational factors significantly affected all the response variables (p<0.05). At Vr > 1 (1.2-3.9), soil reaction forces and drawbar power were considerably reduced. The optimal parameters for minimizing the response variables were 2.5 km h-1 forward speed, 14.5 Hz frequency, 30 mm amplitude, and S-shape shovel at Vr = 3.9. The minimum draught force, vertical force, drawbar power, and Archard wear depth were 4.64 kN, 0.41 kN, 2.64 kW, and 2.36 mm, respectively, at an operating depth of 350 mm. Operating in vibratory mode reduced draught force by 54% with the full width of the implement. Future work should include Jerusalem artichoke tubers in the simulation and experimental validation. Keywords: Abrasive wear, Clay, Numerical optimization, Soil reaction forces, Velocity ratio, Vibration.

Currently, the unmanned rice harvester based on satellite navigation is affected by tracked chassis steering errors and navigation accuracy, which can result in harvesting omissions and undercutting. Therefore, we proposed a cutting width measurement method for unmanned rice harvesters based on RGB-D images. The proposed method can provide navigational aids for driving and may also provide a basis for feed quantity detection and adjustment. We designed a lightweight unharvested area segmentation model (UANet) with a mean pixel accuracy (mPA) and a mean intersection over union (mIoU) of 98.28% and 97.15%, respectively, which are both higher than the lightweight semantic segmentation models, such as DABNet, CFPNet-V2 and ENet. We implemented unharvested area segmentation and harvesting boundary points extraction on RGB images, combined with depth maps to convert the pixel coordinates of harvesting boundary points into 3D world coordinates for cutting width measurement. Through experiments on rice harvest field images under different working conditions, the results showed that the success rate of harvesting boundary points detection was 98.33%, the error of cutting width measurement was within 2.34%, and the speed of the algorithm was approximately 9.7 fps. The proposed method can be adapted to complex harvesting scenarios and has the advantages of low hardware cost, high measurement accuracy, and good robustness, which can be extended to wheat.

In agricultural settings, the unstructured nature of certain production environments, along with the high complexity and inherent risks of production tasks, poses significant challenges to achieving full automation and effective on-site machine control. Remote control technology, which leverages human intelligence and precise machine movements, ensures operator safety and boosts productivity. Recently, virtual reality (VR) has shown promise in remote control applications by overcoming single-view limitations and providing three-dimensional information, yet most studies have not focused on agricultural settings. Therefore, to bridge the gap, this study proposes a large-scale digital mapping and immersive human-machine teleoperation framework specifically designed for precision agriculture. In this research, a DJI unmanned aerial vehicle (UAV) was utilized for data collection, and a novel video segmentation approach based on feature points was introduced. To accommodate the variability of complex textures, this method proposes an enhanced Structure from Motion (SfM) approach. It integrates the open Multiple View Geometry (OpenMVG) framework with Local Features from Transformers (LoFTR). The enhanced SfM produces a point cloud map, which is further processed through Multi-View Stereo (MVS) to generate a complete map model. For control, a closed-loop system utilizing TCP/IP for VR control and positioning of agricultural machinery was introduced. This system offers a fully visual-based method for immersive control, allowing operators to utilize VR technology for remote operations. The experimental results demonstrate that the user-friendly remote control method also showcases its advantages over traditional video streaming-based remote operations, providing operators with a more comprehensive and immersive experience and a higher level of situational awareness.

Rice panicle density is one of the essential bases for the automatic speed regulation of unmanned harvesters, making density detection crucial for intelligent upgrades. Currently, existing methods for detecting rice panicle density do not meet actual harvesting scenarios and struggle to meet real-time requirements. To address this, we developed a real-time rice panicle density detection method for unmanned harvesters. This method includes a panicle detection model based on YOLOv5n (RP-YOLO) and a rice panicle density calculation based on coordinate transformations. RP-YOLO was optimized through various techniques, such as enhancing the target detection head, reconfiguring the backbone network and downsampling module, introducing an attention mechanism, and refining the loss function. Based on coordinate conversion, we converted the world coordinates of the detection frame vertex to image coordinates and calculated the panicle density. We established the RP-1668 dataset for japonica rice and trained and tested the model. Compared to the original YOLOv5n model, our modifications reduced floating-point operations per second (FLOPs) by 33.33 %, decreased model size by 31.90 %, increased detection speed by 12.63 %, and improved accuracy (AP0.5) by 3.82 % (AP0.5:0.95, 6.96 %). RP-YOLO achieved superior accuracy and detection speed compared to both conventional lightweight and non-lightweight models. In field applications, the error in density detection was less than 10 % compared to manual counting, and the results clearly reflected changes in rice panicle density. For a 1.4 m x 1.0 m rice field imaging area (with a resolution of 2560 x 1280), the method detects at 15 fps on an on-board industrial computer, providing reliable data support for adjusting the operating speed of driverless harvesters.