Daniel Queirós Silva

Abstract

Abstract
Object identification, such as tree trunk detection, is fundamental for forest robotics. Intelligent vision systems are of paramount importance in order to improve robotic perception, thus enhancing the autonomy of forest robots. To that purpose, this paper presents three contributions: an open dataset of 5325 annotated forest images; a tree trunk detection Edge AI benchmark between 13 deep learning models evaluated on four edge-devices (CPU, TPU, GPU and VPU); and a tree trunk mapping experiment using an OAK-D as a sensing device. The results showed that YOLOR was the most reliable trunk detector, achieving a maximum F1 score around 90% while maintaining high scores for different confidence levels; in terms of inference time, YOLOv4 Tiny was the fastest model, attaining 1.93 ms on the GPU. YOLOv7 Tiny presented the best trade-off between detection accuracy and speed, with average inference times under 4 ms on the GPU considering different input resolutions and at the same time achieving an F1 score similar to YOLOR. This work will enable the development of advanced artificial vision systems for robotics in forestry monitoring operations.

Abstract
The development of robotic solutions in unstructured environments brings several challenges, mainly in developing safe and reliable navigation solutions. Agricultural environments are particularly unstructured and, therefore, challenging to the implementation of robotics. An example of this is the mountain vineyards, built-in steep slope hills, which are characterized by satellite signal blockage, terrain irregularities, harsh ground inclinations, and others. All of these factors impose the implementation of precise and reliable navigation algorithms, so that robots can operate safely. This work proposes the detection of semantic natural landmarks that are to be used in Simultaneous Localization and Mapping algorithms. Thus, Deep Learning models were trained and deployed to detect vine trunks. As significant contributions, we made available a novel vine trunk dataset, called VineSet, which was constituted by more than 9000 images and respective annotations for each trunk. VineSet was used to train state-of-the-art Single Shot Multibox Detector models. Additionally, we deployed these models in an Edge-AI fashion and achieve high frame rate execution. Finally, an assisted annotation tool was proposed to make the process of dataset building easier and improve models incrementally. The experiments show that our trained models can detect trunks with an Average Precision up to 84.16% and our assisted annotation tool facilitates the annotation process, even in other areas of agriculture, such as orchards and forests. Additional experiments were performed, where the impact of the amount of training data and the comparison between using Transfer Learning and training from scratch were evaluated. In these cases, some theoretical assumptions were verified.

Abstract
Smart and precision agriculture concepts require that the farmer measures all relevant variables in a continuous way and processes this information in order to build better prescription maps and to predict crop yield. These maps feed machinery with variable rate technology to apply the correct amount of products in the right time and place, to improve farm profitability. One of the most relevant information to estimate the farm yield is the Leaf Area Index. Traditionally, this index can be obtained from manual measurements or from aerial imagery: the former is time consuming and the latter requires the use of drones or aerial services. This work presents an optical sensing-based hardware module that can be attached to existing autonomous or guided terrestrial vehicles. During the normal operation, the module collects periodic geo-referenced monocular images and laser data. With that data a suggested processing pipeline, based on open-source software and composed by Structure from Motion, Multi-View Stereo and point cloud registration stages, can extract Leaf Area Index and other crop-related features. Additionally, in this work, a benchmark of software tools is made. The hardware module and pipeline were validated considering real data acquired in two vineyards—Portugal and Italy. A dataset with sensory data collected by the module was made publicly available. Results demonstrated that: the system provides reliable and precise data on the surrounding environment and the pipeline is capable of computing volume and occupancy area from the acquired data.

Abstract
Object recognition represents the ability of a system to identify objects, humans or animals in images. Within this domain, this work presents a comparative analysis among different classification methods aiming at Tactode tile recognition. The covered methods include: (i) machine learning with HOG and SVM; (ii) deep learning with CNNs such as VGG16, VGG19, ResNet152, MobileNetV2, SSD and YOLOv4; (iii) matching of handcrafted features with SIFT, SURF, BRISK and ORB; and (iv) template matching. A dataset was created to train learning-based methods (i and ii), and with respect to the other methods (iii and iv), a template dataset was used. To evaluate the performance of the recognition methods, two test datasets were built: tactode_small and tactode_big, which consisted of 288 and 12,000 images, holding 2784 and 96,000 regions of interest for classification, respectively. SSD and YOLOv4 were the worst methods for their domain, whereas ResNet152 and MobileNetV2 showed that they were strong recognition methods. SURF, ORB and BRISK demonstrated great recognition performance, while SIFT was the worst of this type of method. The methods based on template matching attained reasonable recognition results, falling behind most other methods. The top three methods of this study were: VGG16 with an accuracy of 99.96% and 99.95% for tactode_small and tactode_big, respectively; VGG19 with an accuracy of 99.96% and 99.68% for the same datasets; and HOG and SVM, which reached an accuracy of 99.93% for tactode_small and 99.86% for tactode_big, while at the same time presenting average execution times of 0.323 s and 0.232 s on the respective datasets, being the fastest method overall. This work demonstrated that VGG16 was the best choice for this case study, since it minimised the misclassifications for both test datasets.