Abstract
Developing ground robots for crop monitoring and harvesting in steep slope vineyards is a complex challenge due to two main reasons: harsh condition of the terrain and unstable localization accuracy obtained with Global Navigation Satellite System. In this context, a reliable localization system requires an accurate and redundant information to Global Navigation Satellite System and wheel odometry based system. To pursue this goal we benchmark 3 well known Visual Odometry methods with 2 datasets. Two of these are feature-based Visual Odometry algorithms: Libviso2 and SVO 2.0. The third is an appearance-based Visual Odometry algorithm called DSO. In monocular Visual Odometry, two main problems appear: pure rotations and scale estimation. In this paper, we focus on the first issue. To do so, we propose a Kalman Filter to fuse a single gyroscope with the output pose of monocular Visual Odometry, while estimating gyroscope's bias continuously. In this approach we propose a non-linear noise variation that ensures that bias estimation is not affected by Visual Odometry resultant rotations. We compare and discuss the three unchanged methods and the three methods with the proposed additional Kalman Filter. For tests, two public datasets are used: the Kitti dataset and another built in-house. Results show that our additional Kalman Filter highly improves Visual Odometry performance in rotation movements.

Abstract
Monocular Visual Odometry techniques represent a challenging and appealing research area in robotics navigation field. The use of a single camera to track robot motion is a hardware-cheap solution. In this context, there are few Visual Odometry methods on the literature that estimate robot pose accurately using a single camera without any other source of information. The use of omnidirectional cameras in this field is still not consensual. Many works show that for outdoor environments the use of them does represent an improvement compared with the use of conventional perspective cameras. Besides that, in this work we propose an open-source monocular omnidirectional version of the state-of-the-art method Libviso2 that outperforms the original one even in outdoor scenes. This approach is suitable for central dioptric omnidirectional cameras and takes advantage of their wider field of view to calculate the robot motion with a really positive performance on the context of monocular Visual Odometry. We also propose a novel approach to calculate the scale factor that uses matches between laser measures and 3-D triangulated feature points to do so. The novelty of this work consists in the association of the laser ranges with the features on the omnidirectional image. Results were generate using three open-source datasets built in-house showing that our unified system largely outperforms the original monocular version of Libviso2.

Abstract
Developing ground robots for crop monitoring and harvesting in steep slope vineyards is a complex challenge due to two main reasons: harsh condition of the terrain and unstable localization accuracy obtained with Global Navigation Satellite System. In this context, a reliable localization system requires an accurate and redundant information to Global Navigation Satellite System and wheel odometry based system. To pursue this goal and have a reliable localization system in our robotic platform we aim to extract the better performance as possible from a monocular Visual Odometry method. To do so, we present a benchmark of Libviso2 using both perspective and fisheye lens cameras, studying the behavior of the method using both topologies in terms of motion performance in an outdoor environment. Also we analyze the quality of feature extraction of the method using the two camera systems studying the impact of the field of view and omnidirectional image rectification in VO. We propose a general methodology to incorporate a fisheye lens camera system into a VO method. Finally, we briefly describe the robot setup that was used to generate the results that will be presented. © 2019, Springer Nature Switzerland AG.

Abstract
The main task while developing a mobile robot is to achieve accurate and robust navigation in a given environment. To achieve such a goal, the ability of the robot to localize itself is crucial. In outdoor, namely agricultural environments, this task becomes a real challenge because odometry is not always usable and global navigation satellite systems (GNSS) signals are blocked or significantly degraded. To answer this challenge, this work presents a solution for outdoor localization based on an omnidirectional visual odometry technique fused with a gyroscope and a low cost planar light detection and ranging (LIDAR), that is optimized to run in a low cost graphical processing unit (GPU). This solution, named FAST-FUSION, proposes to the scientific community three core contributions. The first contribution is an extension to the state-of-the-art monocular visual odometry (Libviso2) to work with omnidirectional cameras and single axis gyro to increase the system accuracy. The second contribution, it is an algorithm that considers low cost LIDAR data to estimate the motion scale and solve the limitations of monocular visual odometer systems. Finally, we propose an heterogeneous computing optimization that considers a Raspberry Pi GPU to improve the visual odometry runtime performance in low cost platforms. To test and evaluate FAST-FUSION, we created three open-source datasets in an outdoor environment. Results shows that FAST-FUSION is acceptable to run in real-time in low cost hardware and that outperforms the original Libviso2 approach in terms of time performance and motion estimation accuracy.

Abstract
Agricultural robotics is nowadays a complex, challenging, and exciting research topic. Some agricultural environments present harsh conditions to robotics operability. In the case of steep slope vineyards, there are several challenges: terrain irregularities, characteristics of illumination, and inaccuracy/unavailability of signals emitted by the Global Navigation Satellite System (GNSS). Under these conditions, robotics navigation becomes a challenging task. To perform these tasks safely and accurately, the extraction of reliable features or landmarks from the surrounding environment is crucial. This work intends to solve this issue, performing accurate, cheap, and fast landmark extraction in steep slope vineyard context. To do so, we used a single camera and an Edge Tensor Processing Unit (TPU) provided by Google & x2019;s USB Accelerator as a small, high-performance, and low power unit suitable for image classification, object detection, and semantic segmentation. The proposed approach performs object detection using Deep Learning (DL)-based Neural Network (NN) models on this device to detect vine trunks. To train the models, Transfer Learning (TL) is used on several pre-trained versions of MobileNet V1 and MobileNet V2. A benchmark between the two models and the different pre-trained versions is performed. The models are pre-trained in a built in-house dataset, that is publicly available containing 336 different images with approximately 1,600 annotated vine trunks. There are considered two vineyards, one using camera images with the conventional infrared filter and others with an infrablue filter. Results show that this configuration allows a fast vine trunk detection, with MobileNet V2 being the most accurate retrained detector, achieving an overall Average Precision of 52.98 & x0025;. We briefly compare the proposed approach with the state-of-the-art Tiny YOLO-V3 running on Jetson TX2, showing the outperformance of the adopted system in this work. Additionally, it is also shown that the proposed detectors are suitable for the Localization and Mapping problems.

Abstract
Research and development in mobile robotics are continuously growing. The ability of a human-made machine to navigate safely in a given environment is a challenging task. In agricultural environments, robot navigation can achieve high levels of complexity due to the harsh conditions that they present. Thus, the presence of a reliable map where the robot can localize itself is crucial, and feature extraction becomes a vital step of the navigation process. In this work, the feature extraction issue in the vineyard context is solved using Deep Learning to detect high-level features - the vine trunks. An experimental performance benchmark between two devices is performed: NVIDIA's Jetson Nano and Google's USB Accelerator. Several models were retrained and deployed on both devices, using a Transfer Learning approach. Specifically, MobileNets, Inception, and lite version of You Only Look Once are used to detect vine trunks in real-time. The models were retrained in a built in-house dataset, that is publicly available. The training dataset contains approximately 1600 annotated vine trunks in 336 different images. Results show that NVIDIA's Jetson Nano provides compatibility with a wider variety of Deep Learning architectures, while Google's USB Accelerator is limited to a unique family of architectures to perform object detection. On the other hand, the Google device showed an overall Average precision higher than Jetson Nano, with a better runtime performance. The best result obtained in this work was an average precision of 52.98% with a runtime performance of 23.14 ms per image, for MobileNet-V2. Recent experiments showed that the detectors are suitable for the use in the Localization and Mapping context.