Journal of Field Robotics最新文献

Simultaneous localization and mapping (SLAM) is at the core of robotics automation, relying on sensors such as Laser Detection and Ranging (LiDAR) and cameras to digitally construct a robot's environment and determine its position within. LiDAR-based SLAM outperforms visual-SLAM, especially in low visibility and challenging lighting conditions. However, these systems still face challenges like scene degradation when dealing with feature-deficient degenerate environments such as long corridors or tunnels. Traditional LiDAR SLAM algorithms primarily focus on the extraction of geometric features from the scene, with less utilization of visual information, for example, LiDAR-generated reflectivity (also commonly referred to as intensity image) and depth imagery. In this study, we explore the potential of fusing both geometric and LiDAR-generated image features into the SLAM system in various forms, aiming to enhance the system's adaptability in diverse environments and its robustness against environment degeneracy. We propose a new multifeature-modality SLAM designed for robust real-time localization and mapping in challenging environments. Our method enhances and extracts visual features from LiDAR-generated images, which are then fused with geometric features through a holistic residual function for pose optimization. We also integrate a deep learning-based object removal algorithm to reduce sensitivity to moving objects and sensor noise. This article conducts an in-depth comparison of the proposed algorithm with several leading technologies in terms of scan matching accuracy, robustness, odometry, and mapping. The experimental results vividly showcase the superiority of our method in achieving high scan matching success rates and strong resilience against random outliers and Gaussian noise across various challenging scenarios, compared to the existing LiDAR SLAM methods that rely solely on geometric features. Extensive field experiments conducted on publicly available data sets, along with independently developed backpack-based and robotic platforms, validated the robustness and accuracy of the proposed approach in both indoor and outdoor environments. In 3D mapping, we quantified the precision of 3D points by comparing point clouds collected by high-precision Mobile Laser Scanning (MLS) and Terrestrial Laser Scanning (TLS). Our method outperforms in terms of absolute pose errors (APE) and point cloud matching quality. Based on the fitted Weibull distribution, the root mean square error (RMS) of point-to-plane distances improved by 20%. Additionally, ablation tests revealed the efficacy of different components within our system.

In cotton picking, there is a need for a robotic arm that can pick bolls efficiently with minimal loss and damage to immature bolls. This study addresses that gap by developing a hydraulic arm to work effectively in unstructured cotton fields. A three-degrees-of-freedom hydraulic arm was developed with vertical, radial, and rotational movements, equipped with a roller-type end-effector and vacuum suction system to pick cotton bolls from the plants and collect them into a storage tank. The performance of the hydraulic arm was evaluated in the field under varying conditions of plant-to-plant spacing (250, 350, and 450 mm) and at suction pressures (25, 50, and 75 mmHg). Key performance parameters such as picking efficiency, picking capacity, damage of green bolls, losses of cotton bolls, and trash content were evaluated. The results indicated that the maximum picking efficiency (92.79%) and capacity (567 bolls picked/h) were achieved at 450 mm plant-to-plant spacing and 75 mmHg suction pressure, with minimal damage of green bolls (126 bolls/ha) and losses of cotton bolls (3.79%). The lowest trash content (2.06%) was observed at 450 mm spacing and 25 mmHg suction pressure. Finally, the hydraulic arm achieved precise and accurate cotton picking in complex cotton fields, showing the best performance at 450 mm plant-to-plant spacing and 75 mmHg suction pressure.

This paper proposes a monocular visual-based navigation state estimator designed to operate onboard cost-effective Autonomous Underwater Vehicles (AUVs) in monitoring and inspection applications. The estimator exploits a monocular visual odometry solution, named Mono UVO (Monocular Underwater Visual Odometry), integrating acoustic range information to make the scale observable and provide an estimate of the robot linear velocity in complex and unstructured underwater scenarios. By utilizing an Extended Kalman Filter, the visual-based linear velocity is fused with robot attitude and depth measurements to retrieve the AUV navigation state. The proposed navigation framework was extensively tested offline using heterogeneous data sets of real underwater images collected during several experimental campaigns. Moreover, online validation of the navigation state estimator was performed onboard an AUV to accomplish a closed-loop autonomous survey at sea. The performance of the state estimator is evaluated by comparing the estimation output with reference signals obtained from Doppler Velocity Log measurements and GPS when available. The results demonstrate the feasibility of the proposed visual-based state estimator in providing reliable AUV navigation state in very different and challenging underwater environments. Among the contributions, the source code of the Mono UVO algorithm is made available online, together with the release of an underwater data set.

Complex physicochemical environmental effects result in highly intricate and diverse backgrounds in underwater object images, posing significant challenges for detecting structural defects. Besides, current methods overlook the tradeoff between the detection accuracy and computational cost. To effectively address the aforementioned issues, we present a lightweight multiscale feature fusion network (LMFFNet) for underwater structural defect detection. Aiming to enhance the feature representability, spatial attention, and bidirectional pyramid modules are jointly employed for fusing multiscale features. A parameter-sharing header module is designed to reduce the model parameters. The dynamic nonmonotonic focusing mechanism is introduced in the loss function, which can improve the defect detection performance on degraded underwater images. Comprehensive experiments on a real-world underwater data set demonstrate the superiority of the LMFFNet over existing state-of-the-art methods.

The grape and wine industry suffers substantial losses annually due to diseases like downy mildew and grapevine leafroll-associated virus 3. Effective control of these diseases hinges on precise and timely diagnosis, which is often hindered by the shortage of highly skilled disease scouts. This highlights the urgent need for alternative, scalable solutions. We introduce PhytoPatholoBot (PPB), a fully autonomous ground robot equipped with a custom imaging system and onboard analysis pipeline for near-real-time disease detection and severity quantification, enabling rapid disease assessments in vineyards. The imaging system uses active illumination to enhance image quality and consistency, addressing a key challenge in ensuring the generalizability of analysis models. The analysis pipeline incorporates a disease mapping near-real-time model, a custom segmentation model designed for deployment on low-power edge computing devices, allowing near-real-time inference. PPB was deployed in both research and commercial vineyards for field-based disease scouting. Experimental results demonstrated that its disease detection and severity quantification performance was comparable to those of experienced human scouts and advanced offline computer vision models, while maintaining high computational efficiency and low-power consumption suited to field robots. PPB's ability to map disease progression over the growing season and manage multiple disease types in previously unseen vineyards highlights its potential to advance agricultural research and improve vineyard disease management practices.