Abstract Chain of Thought (CoT) prompting has been shown to improve the performance of large language models (LLMs) in a wide range of tasks, including arithmetic, common-sense, and symbolic reasoning. However, this improvement requires the development of effective CoT prompts. On the other hand, more recent work has shown that CoT reasoning paths are often inherently present in top-k alternative decoding sequences, even in the absence of any specific prompting technique. In this study, we propose a new fine-tuning method that exploits this property by targeting only two specific tokens of these pre-existing CoT responses. We demonstrate that fine-tuning only two tokens using the model’s own implicitly generated CoT paths leads to a significant efficiency gain, reducing training time while still achieving meaningful performance improvements. When evaluated on arithmetic datasets, we achieved a 22.7% improvement on MultiArith, 9.0% on GSM8K, and 2.3% on SVAMP when validated on the Phi-2 model from a greedy decoding perspective, reducing the processing time by over 90% compared to the LoRA fine-tuning method. Code is publicly available at: https://github.com/paulosantosneto/2tft .

The vessels are powered by propellers that convert engine power into movement. Optimizing propeller design involves numerous variables like diameter and blade count, making exact methods impractical. Meta-heuristics, such as evolutionary algorithms, offer a promising solution. This study introduces a novel marine propeller optimization approach. It decomposes the optimization process and proposes a new fitness function to overcome previous limitations. Based on two continuous optimization algorithms, the approach was compared to the state-of-the-art differential evolution algorithm. Results from ferry-boat propeller design experiments show the proposed approach achieves approximately 1% higher efficiency than the baseline study at 7.0 and 7.5-knot speeds. The proposed approach succeeds at 8.0 and 8.5 knots, where the baseline failed. Additionally, decomposition reduces execution time by executing in six threads.

Activities in underwater environments are paramount in several scenarios, which drives the continuous development of underwater image enhancement techniques. A major challenge in this domain is the depth at which images are captured, with increasing depth resulting in a darker environment. Most existing methods for underwater image enhancement focus on noise removal and color adjustment, with few works dedicated to brightness enhancement. This work introduces a novel unsupervised learning approach to underwater image enhancement using a diffusion model. Our method, called UDBE, is based on conditional diffusion to maintain the brightness details of the unpaired input images. The input image is combined with a color map and a Signal-Noise Relation map (SNR) to ensure stable training and prevent color distortion in the output images. The results demonstrate that our approach achieves an impressive accuracy rate in the datasets UIEB, SUIM and RUIE, well-established underwater image benchmarks. Additionally, the experiments validate the robustness of our approach, regarding the image quality metrics PSNR, SSIM, UIQM, and UISM, indicating the good performance of the brightness enhancement process. The source code is available here.

Given the significant increase in the global population, food production difficulties arise. Therefore, using technologies in agricultural production is necessary, especially in spraying, where precision agriculture reduces the damage to the health of workers, consumers, and the environment. In Brazil, there is a problem related to inequality in distributing this high-cost technology to small-scale producers. Thus, this work proposes developing and validating a low-cost solenoid valve actuation system for flow control, with variable rate via Pulse Width Modulation (PWM). In addition, a Simulink model is proposed to simulate the created spraying prototype. We performed a sensitivity analysis to validate the system’s physical parameters. This work carried out an experimental evaluation with four models of solenoid valves, with one commercial valve costing ten times more than the others. The performance of the flow response with frequency variation and duty cycle in each valve is evaluated. The analysis of the results shows that the best performance is obtained when using the normally open type compared to the normally closed type for PWM control applications.

Quality control is a paramount task for the manufacturing process at the current time. Nowadays, industrial companies must guarantee the quality of their products to keep competitiveness. This paper addresses the problem of controlling the quality of GSM chips during the soldering process of individual chips in PCB substrates, providing quality to the manufacturing stage. We propose an approach to detect solder mask faults in GSM chips, during the soldering stage. The proposed methodology is based on the use of Vision Transformers (ViT), which use self-attention mechanisms to capture the spatial relationships between image patches, allowing the model to learn hierarchical visual representations. Real and simulated experiments were carried out to validate the proposed approach. Results show the obtained accuracy of 95.83%, using the proposed ViT-based inspection approach. Furthermore, the proposed approach presents high accuracy even regarding noisy and blurry images, resulting in an accuracy of 94.94% and 94.17% for Salt and Pepper and Gaussian noise, respectively, in the worst scenario. Experiments demonstrate reliability and robustness, optimizing the manufacturing.

Demand for ocean-based high-quality and sustainable fish protein soared in the last decade. Unlike precision agriculture, aquaculture remains an under-equipped farming activity. The aquaculture industry has provided remarkable contributions to the Sustainable Development Goal of zero hunger based on providing animal-based protein for human consumption worldwide. The success of the aquaculture industry hinges on appropriate monitoring of key water quality indicators to ensure both animal health and optimal productivity. In this context, the present work presents a cloud-based LoRaWAN system for quasi-real-time tracking of essential water quality parameters by integrating Internet of Things (IoT) sensor devices. The proposed approach harnesses the power of Long Range (LoRa) technology - especially the LoRa Wide Area Network (LoRaWAN) protocol - to facilitate efficient, large-scale monitoring focusing on data security and scalability. With practical insights drawn from IoT system deployment at an industrially relevant aquaculture farm in Brazil, this research provides a comprehensive look into the system's capabilities, drawbacks, and end-user feedback, offering a blueprint for future aquaculture innovations.

This article reviews computer vision technologies for detecting and tracking objects in industrial cargo handling activities. We have proposed a dataset and a methodology for identifying people, containers, cages, equipment, boxes, and piping, in real-time operation. Our experimental results demonstrate that our artificial neural network model effectively detects and segments objects in non-sparse environments using an annotated industrial image dataset, achieving average precision up to 95% for most classes, including 93% of test instances. This improved perception capability enhances operators' decision-making and accident prevention.

Despite the critical role of fertilizers in modern agriculture, the lack of properly labeled datasets has significantly hindered advancements in automated fertilizer analysis. To address this gap, this paper introduces three novel datasets tailored for the development and validation of fertilizer detection and classification systems. First, a synthetic dataset is generated using a surface simulator that combines images of individual fertilizer grains, providing a highly controlled yet diverse data source for preliminary algorithm testing. Second, a controlled environment dataset is created under optimal yet realistic conditions to offer a balance between controlled experiments and applicability in natural settings. Third, a real-environment dataset is compiled under challenging field conditions, which presents the complexities of real-world agricultural data collection. Together, these datasets not only enhance the training and testing of machine learning models but also pave the way for substantial improvements in precision agriculture by enabling more accurate and efficient fertilizer management. This paper details the creation, characteristics, and potential applications of these datasets, aiming to set a new standard for dataset quality and utility in agricultural research.

Vehicles capable of operating in more than one environment have been developed to solve real problems. Among them, the hybrid unmanned aerial-underwater vehicle (HUAUV) is receiving attention from the robotics community, mainly with a quadrotor-like configuration. However, this vehicle presents high energy consumption because of the larger mass required compared to the only aerial vehicle, limiting its autonomy. This work addresses the trajectory planning problem for a HUAUV. The method is based on Rapidly-exploring Random Trees (RRTs), a highly customizable planning technique. In addition, we propose two new heuristics to increase the energy efficiency of the hybrid vehicle. The first consists of biasing the tree expansion towards the environment with the lowest navigation cost, while the second one assigns estimated costs to nodes in the tree and chooses the least expensive trajectories. These techniques are evaluated in physically realistic simulation experiments performed in 135 scenarios. A comparative analysis of their performances is presented relative to the state of the art. We show that using efficient heuristics can significantly contribute to reducing energy consumption and even increase the average velocity in the missions performed by these vehicles.

Bird's-Eye-View (BEV) semantic maps have become an essential component of automated driving pipelines due to the rich representation they provide for decision-making tasks. However, existing approaches for generating these maps still follow a fully supervised training paradigm and hence rely on large amounts of annotated BEV data. In this work, we address this limitation by proposing the first self-supervised approach for generating a BEV semantic map using a single monocular image from the frontal view (FV). During training, we overcome the need for BEV ground truth annotations by leveraging the more easily available FV semantic annotations of video sequences. Thus, we propose the SkyEye architecture that learns based on two modes of self-supervision, namely, implicit supervision and explicit supervision. Implicit supervision trains the model by enforcing spatial consistency of the scene over time based on FV semantic sequences, while explicit supervision exploits BEV pseudolabels generated from FV semantic annotations and self-supervised depth estimates. Extensive evaluations on the KITTI-360 dataset demonstrate that our self-supervised approach performs on par with the state-of-the-art fully supervised methods and achieves competitive results using only 1 % of direct supervision in BEV compared to fully supervised approaches. Finally, we publicly release both our code and the BEV datasets generated from the KITTI-360 and Waymo datasets.

Fisheries and aquaculture industries notably contribute to animal-source protein production worldwide. Climate change is creating environmental conditions suitable for harmful algal blooms (HAB) on a global scale. Some phytoplankton species can also release toxins, which may cause large-scale marine mortality with knock-on effects on coastal economies. Reliable phytoplankton monitoring and early HAB detection are also essential in climate-resilient solutions for aquaculture applications. Currently, phytoplankton monitoring is primarily based on traditional microscopy. However, it is time-consuming and requires an experienced taxonomist. There is a need to expedite and automate phytoplankton monitoring to support aquaculture industries. Analytical instruments based on microscopy coupled with artificial intelligence (AI) models may be vital to monitoring applications. Digital plankton data sets are usually imbalanced and reflect natural environmental differences. The lack of data to represent minority species/genera prevents AI models from understanding some taxa completely. It compromises system reliability for HAB monitoring applications. The present study investigates state-of-the-art models for class imbalance problems tailored for HAB monitoring within multi-trophic aquaculture farms from Brazil, South Africa, and Scotland. A unified benchmark database covering publicly available microscopic image-based datasets supported phytoplankton modelling. AI deep collaborative models and threshold moving techniques provided the best results compared to standard architectures. It prevailed, especially for low-abundant yet toxic organisms.

The corrosion inspection process in ship tanks used by the oil industry for the production, storage, and disposal of oil, which is known as Floating Production Storage and Offloading (FPSO), is predominantly manual. It requires a long production downtime, and is an unhealthy job for inspectors. In the literature, some works proposed methods for corrosion segmentation. However, none of them classifies the level of corrosion in accordance with the International Association of Classification Societies (IACS) standard. This work proposes the use of U-Net-based network for segmentation of pitting corrosion, and also provides a corrosion level analysis algorithm relating the identified pitting to the IACS standard. Furthermore, data augmentation methods are adopted to make the dataset more diversified, aiming to generalize the neural network learning. The results indicate a mean squared error of only 0.1639 using the proposed method, and an intersection-of-union of 0.9453. In addition, we compared our method with classical methods such as Canny, Laplacian, Otsu, and Sobel methods, where a relevant advantage is obtained with U-Net.

Uncertainty estimation is crucial in safety-critical settings such as automated driving as it provides valuable information for several downstream tasks including high-level decision making and path planning. In this work, we propose EvCenterNet, a novel uncertainty-aware 2D object detection framework using evidential learning to directly estimate both classification and regression uncertainties. To employ evidential learning for object detection, we devise a combination of evidential and focal loss functions for the sparse heatmap inputs. We introduce class-balanced weighting for regression and heatmap prediction to tackle the class imbalance encountered by evidential learning. Moreover, we propose a learning scheme to actively utilize the predicted heatmap uncertainties to improve the detection performance by focusing on the most uncertain points. We train our model on the KITTI dataset and evaluate it on challenging out-of-distribution datasets including BDD100K and nuImages. Our experiments demonstrate that our approach improves the precision and minimizes the execution time loss in relation to the base model.

The present work presents an artificial neural network architecture for the restoration of images damaged by underexposure and overexposure. The problem is relevant in computer vision applications that are applied in conditions where the limitation of the sensor prevent the scene details from being adequately represented in the captured image. This research presents an attention-based architecture composed of two convolutional neural networks, where one performs a preprocessing of the input image, while the other performs the restoration and enhancement of the degraded image. Regarding the evaluation of research results, a broad range of image quality metrics is used to assess the quality of the results produced by the model. The obtained results indicate that the proposed architecture is able to enhance images damaged by exposure heterogeneity, offering gains over state-of-art models in real data.

Fertilizers are fundamental compounds to balance nutrients in the soil, ensuring its fertility for food production. In the industry of fertilizers, a common task is the overhaul of the pipelines that convey the material through production lines, which need to be performed periodically, to avoid duct blockages. Traditionally, this task is carried out manually, which requires interruption of production. Therefore, it implies time consumption and waste of money, in the case of unnecessary inspection. To avoid needless production stoppage, in this paper is presented a non-invasive overhaul method for sediment detection in the pipelines of fertilizer production lines based in neural networks. The proposed model uses thermal images to estimate the volume of sediments into pipelines. Furthermore, as it is difficult to obtain images of several pipeline blockage conditions, a methodology for artificial dataset creation is also provided. The results indicate the feasibility of the proposed methodology.

The process of size analysis of grains in the fertilizer industry is slow, because it is performed by sieves. As an alternative to this mechanized process, digital image techniques have been used to segment and analyze particles in the quality analysis of the grains. However, most deterministic methods for image segmentation do not present high performance when there is no background in the scene, which provides the contrast with the object to be segmented. Furthermore, these methods only ensure its accuracy for segmentation of the objects class considered in the algorithm calibration. Therefore, taking into account this constraint and the great variety of grain size in the fertilizer production process, this paper proposes to use a neural network, U-net, for generalization of grain segmentation, considering a fully covered surface scene, where there is no background. Besides, to show the advantages of proposed solution, a comparison of neural network with deterministic methods is also provided.

Reinforcement Learning (RL) has presented an impressive performance in video games through raw pixel imaging and continuous control tasks. However, RL performs poorly with high-dimensional observations such as raw pixel images. It is generally accepted that physical state-based RL policies such as laser sensor measurements give a more sample-efficient result than learning by pixels. This work presents a new approach that extracts information from a depth map estimation to teach an RL agent to perform the mapless navigation of Unmanned Aerial Vehicle (UAV). We propose the Depth-Imaged Contrastive Unsupervised Prioritized Representations in Reinforcement Learning (Depth-CUPRL) that estimates the depth of images with a prioritized replay memory. We used a combination of RL and Contrastive Learning to lead with the problem of RL based on images. From the analysis of the results with Unmanned Aerial Vehicles (UAVs), it is possible to conclude that our Depth-CUPRL approach is effective for the decision-making and outperforms state-of-the-art pixel-based approaches in the mapless navigation capability.

Since the application of Deep Q-Learning to the continuous action domain in Atari-like games, Deep Reinforcement Learning (Deep-RL) techniques for motion control have been qualitatively enhanced. Nowadays, modern Deep-RL can be successfully applied to solve a wide range of complex decision-making tasks for many types of vehicles. Based on this context, in this paper, we propose the use of Deep-RL to perform autonomous mapless navigation for Hybrid Unmanned Aerial Underwater Vehicles (HUAUVs), robots that can operate in both, air or water media. We developed two approaches, one deterministic and the other stochastic. Our system uses the relative localization of the vehicle and simple sparse range data to train the network. We compared our approaches with an adapted version of the BUG2 algorithm for mapless navigation of aerial vehicles. Based on experimental results, we can conclude that Deep-RL-based approaches can be successfully used to perform mapless navigation and obstacle avoidance for HUAUVs. Our vehicle accomplished the navigation in two scenarios, being capable to achieve the desired target through both environments, and even outperforming the behavior-based algorithm on the obstacle-avoidance capability.

The production of food for all the population in the world became the biggest concern. The population continues to grow and the number of farmable lands has been decreasing. To make the lands more productive, fertilizers are used on a larger scale. To guarantee the quality of the product, particle size analysis are made by mechanical sieving. With the time, the wear-out of the sieving in the fertilizer industry the results of the particle size analysis will be erroneous. So the computer vision appears as an alternative that is non-invasive and less time-consuming. In this context, this paper has the objective to develop a grain surface simulator capable of generating virtual images with overlapping grains, since there is a difficulty to obtain annotated data of images of fertilizers. In order to validate the proposed simulator using a DIP algorithm, noises are added in the virtual images to compare with the reality in the industry, to show how well the particle size analysis with computer vision were handled towards adversities. The results of the overlapping analysis show that when the virtual image has a fewer number of grains, the DIP algorithm can identify the majority of grains, consequently with less error in the particle size analysis. Different noises, at different intensities, have their effects analyzed on the algorithm. As the analyzes in this study match with the reality showing the consequences, tendencies, and errors of the overlapping of grains and noises in the images, the simulator developed here matches with reality and is extremely useful to facilitate the study of complex cases of application of visual computing and digital image processing in particle size analysis of fertilizers.

The continuous growth of the naval and offshore industries require that the processes be steady and reliable, forcing the industries to keep up and dealing with the most modern methods of development. Among these processes, the welding is one of the most commons that has been used widely in the construction of oil platforms. The welding is used in many areas, beyond the naval, it being dangerous for the welder because the surroundings contain fumes and radiation. Once the welding is an usual procedure and the surroundings of the work space can be harmful to human's health, bring more automation for the welding process can improve the productivity levels and move human operators away from danger. However in the naval industry the projects are individualized, making the process of implementation of robotic cells difficult. Using modular robots turn the weld more reliable but are not a optimal solution due it still need a human operator near the process to change the robot's parameters and detect possible errors. So, this work proposes a architecture to turn these robots into autonomous systems, being able to detect the features of the groove to be weld, perform the welding with minimum of human interference or none and inspect the final product.

Image restoration and image enhancement are critical image processing tasks since good image quality is mandatory for many image applications. We are particularly interested in the restoration of ill-exposed images. These effects are caused by sensor limitation or optical arrangement. They prevent the details of the scene from being adequately represented in the captured image. We proposed a deep neural network model due to the number of uncontrolled variables that impact the acquisition. The proposed network can converge in a representative model from the training data, loss, optimization and activation functions. The obtained results are evaluated using several image quality index which indicate that the proposed network is able to improve images damaged by heterogeneous exposure. Furthermore, our method offers a significant gain over the state-of-the-art methods both in simulated data and real data.

Digital cameras work through transforming the scene's radiance into an electrical charge. Optical arrangement, sensors, and embedded electronics often limit the accuracy of the representation. Scenes with a dynamic range above the capability of the camera or poor lighting are challenging conditions, which usually result in low contrast images. Soft clipping is usually compensated by transforming the power and shifting the image's histogram. However, under extreme conditions, ill exposure results in severe clipping that requires interpolation and painting. We introduce a model of convolutional neural network to perform signal reconstruction and interpolation. It is designed to be used on sRGB images. The results are evaluated using several metrics of image quality that indicate that the proposed network can improve images that are damaged by different conditions of exposure. In addition, our method offers a substantial gain over state-of-the-art methods.

Convolutional neural networks (CNNs) are powerful machine learning models that have become the state of the art in several problems in the areas of computer vision and image processing. Nevertheless, the knowledge of why and how these models present an impressive performance is still limited. There are visualization techniques that can help us to understand the inner working of neural networks. However, they have mostly been applied to classification models. In this paper, we evaluate the application of visualization methods to networks where the input and output are images of proportional dimensions. The results show that visualization brings visual cues associated with how these systems work, helping in their understanding and improvement. We use the knowledge obtained from the visualization of an image restoration CNN to improve the architecture's efficiency with no significant degradation of its performance.

Sonar images pose hindrances when being elucidated for applications such as underwater navigation and localization. On the other hand, satellite images are simpler to be interpreted, but require GPS that is unavailable underwater due to absorption phenomena. Thus, we propose a neural network capable of translating an acoustic image acquired underwater to a textured image. We called the process sonar-to-satellite translation. We adopted a state-of-the-art neural architecture on a dataset comprised of sonar data and their respective satellite images. The experimental results show our method can extract interesting features from acoustic images and generate an informative texture image.

Forward-looking sonars (FLS) are perception sensors that are not affected by underwater turbidity. FLS are used in Remotely Operated Vehicles (ROVs) to help them in the tasks of exploration, navigation and region mapping. Besides the advantages of working with acoustic images rather than optical images, the former presents various challenges inherent to their construction. Classic Computer Vision (CV) algorithms do not achieve the same success with acoustic images. Furthermore, data-driven approaches are dictating the state-of-the-art in several tasks that require feature extraction. For example, Convolutional Neural Networks (CNNs) are already been used in several CV problems such as classification, image matching, image retrieval, place recognition and one-shot learning. CNNs are showing promising results for problems with FLS images as well. Unfortunately, there are as not as many public datasets and methods for FLS problems as we have for optical images. Knowing that CNNs are capable of mapping correctly millions of images into thousands of labels, we are proposing a novel framework of feature learning strategy for FLS images. In order to evaluate how well the methods generalize, we selected three different FLS annotated datasets for our experiments. Two of them are real-world FLS images from a harbour environment from different locations. The third is generated from a custom 3D scene integrated with open-source underwater robot simulators. In our experiments, we compared our method with state-of-the-art approaches in an unknown environment achieving superior results.

In this paper, a study on nonlinear state estimation methods for Hybrid Unmanned Aerial Underwater Vehicles (HUAUVs) is presented. Based on a detailed dynamic model simulation, we analyse and elect the best nonlinear algorithm among those presented in the state-of-the-art literature addressing local derivative-free nonlinear Kalman Filters (KFs): the Unscented Kalman Filter (UKF), the Cubature Kalman Filter (CKF) and the Transformed Unscented Kalman Filter (TUKF). Here, these three nonlinear probabilistic estimators were compared in terms of the Root Mean Square Error (RMSE) and the average execution time over Monte Carlo simulations. We simulated real-world conditions for our in-production HUAUV prototype using Inertial Measurement Unit (IMU) data and state augmentation for sensor data filtering and trajectory estimation. We have concluded that the CKF proved to be the most interesting KF to low-cost on-board applications for high dimensional state spaces.

Microalgae are unicellular organisms that presents limited physical characteristics such as size, shape or even the present structures. Classifying them manually may require great effort from experts since thousands of microalgae can be found in a small sample of water. Furthermore, the manual classification is a non-trivial operation. We proposed a deep learning technique to solve the problem. We also created a classified dataset that allow us to adopt this technique. To the best of our knowledge, the present work is the first one to apply this kind of technique on the microalgae classification task. The obtained results show the capabilities of the method to properly classify the data by using as input the low resolution images acquired by a particle analyzer instead of pre-processed features. We also show the improvement provided by the use of data augmentation technique.

In recent years, Convolutional Neural Networks (CNNs) have achieved state-of-the-art performance in many image restoration applications. The knowledge of how these models work, however, is still limited. While there have been many attempts at better understanding the inner working of CNNs, they have mostly been applied to classification networks. Because of this, most existing CNN visualization techniques may be inadequate to the study of image restoration architectures. In the paper, we present network inversion, a new method developed specifically to help in the understanding of image restoration Convolutional Neural Networks. We apply our method to underwater image restoration and dehazing CNNs, showing how it can help in the understanding and improvement of these models.

Optical images display drastically reduced visibility due to underwater turbidity conditions. Sonar imaging presents an alternative form of environment perception for underwater vehicles navigation, mapping and localization. In this work we present a novel method for Acoustic Scene Matching. Therefore, we developed and trained a new Deep Learning architecture designed to compare two acoustic images and decide if they correspond to the same underwater scene. The network is named Sonar Matching Network (SMNet). The acoustic images used in this paper were obtained by a Forward Looking Sonar during a Remotely Operated Vehicle (ROV) mission. A Geographic Positioning System provided the ROV position for the ground truth score which is used in the learning process of our network. The proposed method uses 36.000 samples of real data for validation. From a binary classification perspective, our method achieved 98% of accuracy when two given scenes have more than ten percent of intersection.

Welding is an important process in the industrial scenario, especially in the shipbuilding industry. This process is recognized by the laborious work and the hazardous work environment. The use of robots to automate the welding process can reduce the human interference and improve the productivity. This paper proposes a system for automated seam tracking based on passive monocular vision. The vision provides a data feedback to the automated robotic welding system allowing quality and productivity gains. A trajectory controller is developed to correct the robot's movement over the seam reference. The controller and a visual algorithm to find the seam reference in a real Gas Metal Arc Welding (GMAW) process are presented. The proposed system allows the automated seam tracking, the trajectory control of the welding torch, and a higher automation level in linear robotic welding. The capabilities of the method is evaluated using a commercial linear welding robot showing its viability.

For underwater robotics applications involving monitoring and inspection tasks, it is important to capture quality color images in real time. In this paper, we propose a statistically learning method with an automatic selection of the training set for restoring the color of underwater images. Our statistical model is a Markov Random Field with Belief Propagation (MRF-BP). The quality of the results depends strongly on the trained correlations between the degraded image and its corresponding color image. However, it is not possible to have color ground truth data given the inherent conditions of underwater environments. Thus, we build a color adaptive training set by applying a multiple color space analysis to those frames that present a high change in its distribution from the previous frame and use only those frames for training. Experimental results in real underwater video sequences demonstrate that our approach is feasible, even when visibility conditions are poor, as our method can recover and discriminate between different colors in objects that may seem similar to the human eye.

In this paper we propose a new vision-based obstacle avoidance strategy using the Underwater Dark Channel Prior (UDCP) that can be applied to any Unmanned Underwater Vehicle (UUV) equipped with a simple monocular camera and minimal on-board processing capabilities. For each incoming image, our method first computes a relative depth map to estimate the obstacles nearby. Then, the map is segmented and the most promising Region of Interest (RoI) is identified. Finally, an escape direction is computed within the RoI and a control action is performed accordingly to avoid the obstacles. We tested our approach on a video sequence in a natural environment and compared it against a state-of-the-art method showing better performance, specially in light changing conditions. We also provide online results on a low-cost Remotely Operated Vehicle (ROV) in a controlled environment.

This paper presents a method for stabilizing the attitude of a Hybrid Unmanned Aerial Underwater Vehicle. Firstly, we present aerodynamic and hydrodynamic models for the angular motion of our robot, discussing effects like buoyancy force and added inertia. Next, we apply robust control techniques for both environment, aerial and underwater, based on linear uncertain models with only four vertices and well-defined stability criteria, such as D-stability and ℋ2performance. Gain matrices Kairand Kwatare computed and the attitude of the vehicle at the hovering operation point for each environment is controlled, respectively. Finally, a procedure is proposed to check the global stability for the switching control case, when the robot changes from air to water (or vice-versa). Numerical simulations with disturbances and switching control are presented to show the stability at different initial conditions.

Underwater environments present a considerable challenge for computer vision, since water is a scattering medium with substantial light absorption characteristics which is made even more severe by turbidity. This poses significant problems for visual underwater navigation, object detection, tracking and recognition. Previous works tackle the problem by using unreliable priors or expensive and complex devices. This paper adopts a physical underwater light attenuation model which is used to enhance the quality of images and enable the applicability of traditional computer vision techniques images acquired from underwater scenes. The proposed method simultaneously estimates the attenuation parameter of the medium and the depth map of the scene to compute the image irradiance thus reducing the effect of the medium in the images. Our approach is based on a novel optical flow method, which is capable of dealing with scattering media, and a new technique that robustly estimates the medium parameters. Combined with structure-from-motion techniques, the depth map is estimated and a model-based restoration is performed. The method was tested both with simulated and real sequences of images. The experimental images were acquired with a camera mounted on a Remotely Operated Vehicle (ROV) navigating in a naturally lit, shallow seawater. The results show that the proposed technique allows for substantial restoration of the images, thereby improving the ability to identify and match features, which in turn is an essential step for other computer vision algorithms such as object detection and tracking, and autonomous navigation.

This paper proposes a novel methodology to estimate the optical flow in scattering media, which consists on new formulation based on the classical Horn-Schunk approach and the optical image formation model. Our formulation is able to deal with the hard problem of tracking points in a medium where there is absorption and scattering effects. This approach generalizes assumptions of the Horn-Schunk model in order to tackle both non-scattering and scattering media. Our approach uses the Dark Channel Prior to estimate the scene transmission, which attains a significant improvement in the optical flow estimation in scattering media. We show that our approach outperformed state-of-the-art models and we provide a detailed analysis of our technique that shows its applicability to image sequences acquired both in simulated and real scenes.

The complete modeling and simulation of an unmanned vehicle with combined aerial and underwater capabilities, called Hybrid Unmanned Aerial Underwater Vehicle (HUAUV), is presented in this paper. The best architecture for this kind of vehicle was evaluated based on the adaptation of typical platforms for aerial and underwater vehicles, to allow the navigation in both environments. The model selected was based on a quadrotor-like aerial platform, adapted to dive and move underwater. Kinematic and dynamic models are presented here, and the parameters for a small dimension prototype was estimated and simulated. Finally, controllers were used and validated in realistic simulation, including air and water navigation, and the environment transition problem. To the best of our knowledge, it is the first vehicle that is able to navigate in both environment without mechanical adaptation during the medium transitions.

The ability to detect changes is an essential competence that robots should possess for increased autonomy. In several applications, such as surveillance, a robot needs to detect relevant changes in the environment by comparing current sensory data with previously acquired information from the environment. We present an efficient method for point cloud comparison and change detection in 3D environments based on spatial density patterns. Our method automatically segments 3D data corrupted by noise and outliers into an implicit volume bounded by a surface, making it possible to efficiently apply Boolean operations in order to detect changes and to update existing maps. The method has been validated on several trials using mobile robots operating in real environments and its performance was compared to state-of-the-art algorithms. Our results demonstrate the performance of the proposed method, both in greater accuracy and reduced computational cost.

Nowadays, the advance of the technology allows robots to acquire dense point clouds decreasing the price and increasing the performance. However, it is a hard task to deal with due to the large amount of points, the redundancy and the noise. This paper proposes an adaptable system to build a 3D feature model of point clouds using Gaussian Mixture Models. These 3D models are used in order to detect changes in the autonomous robot's working environment. The presented work describes an efficient change detection system based on two consecutive stages. First, a top-down approach estimates features using Gaussian Mixture Models. The presented new approach improves the performance of previous related works in terms of computational load and robustness, nevertheless the system is selection criteria dependent. Thus, the efficiency of different selection criteria are evaluated and compared in this paper. Experimental results demonstrate that the Minimum Distance Length (MDL) criteria outperforms the other studied methods. In the second stage, a change detection method is performed using the previously estimate Mixture of Gaussians. The proposed full system is able to detect changes using Gaussian Mixture Models with a reduced computational cost in relation to state-of-art algorithms.

The ability to detect changes in the environment is an essential trait for robots commissioned to work in several applications. In surveillance, for instance, a robot needs to detect meaningful changes in the environment which is achieved by comparing current sensory data with previously acquired information from the environment. The large amount of sensory data, which are often complex and very noisy, explains the inherent difficulty of this task. As an attempt to tackle this hard problem, we present an efficient method to automatically segment 3D data, corrupted with noise and outliers, into an implicit volume bounded by a surface. The method makes it possible to efficiently apply Boolean operations to 3D data in order to detect changes and to update existing maps. We show that our approach is powerful, albeit simple, with linear time complexity. The method has been validated through several trials using mobile robots operating in real environments and their performance was compared to another state-of-art algorithm. Experimental results demonstrate the performance of the proposed method, both in accuracy and computational cost.

There are several applications for which it is important to both detect and communicate changes in data models. For instance, in some mobile robotics applications (e.g. surveillance) a robot needs to detect significant changes in the environment (e.g. a layout change) which it may achieve by comparing current data provided by its sensors with previously acquired data (e.g. map) of the environment. This often constitutes an extremely challenging task due to the large amounts of data that must be compared in real-time. This paper proposes a framework to detect, and represent changes through a compact model. The main steps of the procedure are: multi-scale sampling to reduce the computation burden; change detection based on Gaussian mixture models; fitting superquadrics to detected changes; and refinement and optimization using the split and merge paradigm. Experimental results in various real and simulated scenarios demonstrate the approach's feasibility and robustness with large datasets.

The ability to detect perceptions which were never experienced before, i.e. novelty detection, is an important component of autonomous robots working in real environments. It is achieved by comparing current data provided by its sensors with a previously known map of the environment. This often constitutes an extremely challenging task due to the large amounts of data that must be compared in real-time. With respect to previously proposed approaches, this paper detects changes in 3D environment based on probabilistic models, the Gaussian Mixture Model, and a fast and robust combined constraint matching algorithm. The matching allows to represent the scene view as a graph which emerges from the comparison between Mixtures of Gaussians. Finding the largest set of mutually consistent matches is equivalent to find the maximum clique on a graph. The proposed approach has been tested for mobile robotics purposes in real environments and compared to other matching algorithms. Experimental results demonstrate the performance of the proposal.

This paper describes an efficient method for retrieving the 3-dimensional shape associated to novelties in the environment of an autonomous robot, which is equipped with a laser range finder. First, changes are detected over the point clouds using a combination of theGaussian mixture model(GMM) and theearth mover's distance(EMD) algorithms. Next, the shape retrieval is achieved using two different algorithms. First, new samplings are generated from each Gaussian function, followed by arandom sampling consensus(RANSAC) algorithm to retrieve geometric primitives. Furthermore, a new algorithm is developed to directly retrieve the shape according to the mathematical space of Gaussian mixture. In this paper, the set of geometric primitives has been limited to the setC = {sphere, cylinder, plane}. The two shape retrieval methods are compared in terms of computational cost and accuracy. Experimental results in various real and simulated scenarios demonstrate the feasibility of the approach.