Deploying deep neural networks on resource-constrained hardware demands aggressive quantization, yet the rounding scheme applied during this process is almost universally chosen for implementation convenience rather than optimized for accuracy. We introduce RoundGrad, a method that treats weight rounding as a first-order optimization problem. Starting from a pretrained model with frozen weights, RoundGrad maintains a learnable score for each weight that determines whether the value rounds up or down; these scores are updated jointly across the entire network via standard backpropagation under a regularized task-loss objective, without layer-wise reconstruction or second-order approximations. Unlike calibration-based post-training quantization methods, RoundGrad optimizes on labeled training data, placing it between traditional PTQ and full quantization-aware training. On ImageNet classification, RoundGrad achieves near-lossless 4-bit weight quantization across standard CNN architectures, using substantially fewer gradient steps than quantization-aware training (up to fewer than standard QAT schedules). Results on Pascal VOC semantic segmentation confirm that the learned rounding policy transfers to dense prediction tasks. The simplicity of the formulation and its competitive accuracy at the deployment-critical 4-bit regime make RoundGrad a practical and principled alternative to existing rounding approaches.

• Polar mappings and a lightweight ViT are used for rotation-invariant learning. • A polar-based position embedding is proposed to enhance the encoding of polar images. • P-RoPE outperforms original RoPE on rotation-invariant fall detection. Rotation-invariant frameworks are crucial in many computer vision tasks, such as human action recognition (HAR), especially when applied in real-world scenarios. Since most datasets, including those on fall detection, have been generated in controlled environments with fixed camera angles, heights, and movements, approaches developed to address such tasks tend to fail when individual appearance variations occur. To address this challenge, our study proposes the use of the EVA-02-Ti lightweight vision transformer for processing people’s polar mappings and handling the task of fall detection. In particular, we strive to leverage the transformation’s rotation-invariant characteristic and correctly classify the rotated images. Towards this goal, a polar-based rotary position embedding (P-RoPE), which generates relative positions among polar patches according to r and θ axes instead of the Cartesian x and y axes, is presented. Replacing the original RoPE, we achieve an enhancement of ViT’s performance, as demonstrated in our experimental protocol, while it also outperforms a state-of-the-art approach. An evaluation was conducted on E-FPDS and VFP290k, where training was performed on initial images and testing was performed on the rotated ones. Finally, when assessed on Fashion-MNIST-rot-12k, a standard dataset for rotation-invariant scenarios, P-RoPE again surpasses both the baseline version and another benchmark method.

Interest in, and the need for, remote learning tools and techniques, as well as the adoption and deployment of digital technologies for tutoring and training, continues to rise. Concurrently, over the past decade, the use of Virtual Reality (VR) applications as a supplementary tool in the learning process has increased significantly. Therefore, VR simulation applications are expected to play a key role in people's education and training in the coming years. In this study, we designed and developed a VR training application called ViRtus -a VR-based system following constructivist and serious game human-computer interaction principles for VR-aided training- and applied it to a real-world scenario, i.e., the construction of an industrial electrical control panel. To demonstrate the potential and sustainability of the VR training system as an alternative to traditional apprenticeship training procedures, we conducted an experimental study comparing the real-world outcomes of three groups trained using a conventional method, a VR-only training, and a hybrid approach combining both. Based on qualitative assessments and the statistical analysis of practical experiments, the collected data and observations indicate that the hybrid VR constructivist training strategy -supplementing conventional trainer instructions with the VR application- can enhance training effectiveness and promote a more productive and sustainable workplace, for both trainees and trainers, in high-risk industrial tasks.

Landing safety is a challenge heavily engaging the research community recently, due to the increasing interest in applications availed by aerial vehicles. In this paper, we propose a landing safety pipeline based on state of the art object detectors and OctoMap. First, a point cloud of surface obstacles is generated, which is then inserted in an OctoMap. The unoccupied areas are identified, thus resulting to a sum of safe landing points. Due to the low processing time achieved by state of the art object detectors and the efficient point cloud manipulation using OctoMap, it is feasible for our approach to deploy on low-weight embedded systems. The proposed pipeline has been evaluated in many simulation scenarios, varying in people density, number, and movement. Simulations were executed with an Nvidia Jetson Nano in the loop to confirm the pipeline’s performance and robustness in a low computing power hardware. The experiments yielded promising results with a 87% success rate.

This study examines the effects of climate change on Kosynthos and Laspias river basins in Eastern Macedonia and Thrace, Greece, focusing on flooding, pollution, and droughts. It uses Earth Observation and stakeholder engagement to analyse climate impact and propose Nature- based, sustainable mitigation and adaptation strategies. These strategies will promote ecosystem and community resilience, emphasising on the need for integrated approaches in urban planning and environmental management for sustainable, resilient aquatic environments.

Reliable and robust navigation of autonomous mobile robots in indoor environments faces significant challenges due to the absence of GPS, visual degradation, repetitive structures, illumination variations, and low texture. These factors adversely affect localization systems. Current robots often use a uniform navigation approach, regardless of the varying localization uncertainties within different indoor environments. In this paper, we propose a holistic, active vision-based path planning method that produces efficient trajectories, aiming to minimize localization error and enhance navigation performance. Specifically, we utilize a 3D model of an indoor environment to derive an Artificial Potential Field (APF) with its associated localizability scores that encapsulate both visual features’ richness and fiducial markers’ placement. APF is employed to direct a Kinematically Constrained Bi-directional Rapidly Exploring Random Tree (KB-RRT) planner towards the calculation of optimal paths, prioritizing high localization areas. Subsequently, we use an online weight-adaptive MPC-based approach that, apart from robust path planning and obstacle avoidance, guides the robot towards areas with the most robust visual features in order to further refine the localization error. The proposed framework has been extensively tested in both simulation and real-world experiments with a mobile robot in a visually challenging indoor environment.

Autonomous vehicle navigation in complex and unpredictable outdoor environments requires extensive and detailed understanding of the surrounding area and compliance with the traffic rules. In this paper, we attempt to imitate human driver behavior towards autonomous navigation that is suitable for diverse, challenging environments, whether urban, semi-structured or rural-like. Our approach starts with a novel method that we propose for extracting free space area using RGB and LiDAR data, in combination with a rough topometric map for route planning. Local goals are extracted in the final drivable region, and the vehicle draws a local path in the free space that is approximately in line with the overall path via a lattice planner. Our method is evaluated both in the publicly available KITTI urban dataset and a custom-made dataset of a semi-structured environment. In both cases, the results highlight the potential of our approach for further advancements in autonomous navigation and the development of safer and more human-like behaviors in driverless vehicles compared to the existing trajectory prediction state-of-the-art methods that make use of a topometric map.

Semantic interpretation of regions or entities is increasingly attracting the attention of scholars, owing to its vast applicability in several disciplines. In this context, modern autonomous systems are capable to semantically recognize and separate entities from camera measurements, while effectively interprete and interact with their environment in a higher level. Extending this notion, the semantic representation of the surroundings, based on satellite and ground-level data, is considered a fundamental property for self-localization, especially in the absence of any georeferencing signal. Keeping that in mind, in this article, we present a robust algorithm to locate the position of an autonomous vehicle within a georeferenced map using graph-based descriptors with semantic and metric information from both its memory and query measurements. In particular, an enhanced prerecorded satellite map is processed to compute semantic memories, whilst ground-level query views are used as a means to identify similarities and extrapolate the location of a moving vehicle. The above components are evaluated under an extensive set of experiments, revealing the robustness and accuracy of our final robot localization system.

The exploitation of deep neural networks (DNNs) as descriptors in feature learning challenges enjoys apparent popularity over the past few years. The above tendency focuses on the development of effective loss functions that ensure both high feature discrimination among different classes, as well as low geodesic distance between the feature vectors of a given class. The vast majority of the contemporary works rely their formulation on an empirical assumption about the feature space of a network's last hidden layer, claiming that the weight vector of a class accounts for its geometrical center in the studied space. This article at hand follows a theoretical approach and indicates that the aforementioned hypothesis is not exclusively met. This fact raises stability issues regarding the training procedure of a DNN, as shown in our experimental study. Consequently, a specific symmetry is proposed and studied both analytically and empirically that satisfies the above assumption, addressing the established convergence issues. More specifically, the aforementioned symmetry suggests that all weight vectors are unit, coplanar, and their vector summation equals zero. Such a layout is proven to ensure a more stable learning curve compared against the corresponding ones succeeded by popular models in the field of feature learning.

One of the most prominent attributes of Neural Networks (NNs) constitutes their capability of learning to extract robust and descriptive features from high dimensional data, like images. Hence, such an ability renders their exploitation as feature extractors particularly frequent in an abundance of modern reasoning systems. Their application scope mainly includes complex cascade tasks, like multi-modal recognition and deep Reinforcement Learning (RL). However, NNs induce implicit biases that are difficult to avoid or to deal with and are not met in traditional image descriptors. Moreover, the lack of knowledge for describing the intra-layer properties -and thus their general behavior- restricts the further applicability of the extracted features. With the paper at hand, a novel way of visualizing and understanding the vector space before the NNs' output layer is presented, aiming to enlighten the deep feature vectors' properties under classification tasks. Main attention is paid to the nature of overfitting in the feature space and its adverse effect on further exploitation. We present the findings that can be derived from our model's formulation and we evaluate them on realistic recognition scenarios, proving its prominence by improving the obtained results.

The advancement of Human-Robot Interaction (HRI) drives research into the development of advanced emotion identification architectures that fathom audio-visual (A-V) modalities of human emotion. State-of-the-art methods in multi-modal emotion recognition mainly focus on the classification of complete video sequences, leading to systems with no online potentialities. Such techniques are capable of predicting emotions only when the videos are concluded, thus restricting their applicability in practical scenarios. This article provides a novel paradigm for online emotion classification, which exploits both audio and visual modalities and produces a responsive prediction when the system is confident enough. We propose two deep Convolutional Neural Network (CNN) models for extracting emotion features, one for each modality, and a Deep Neural Network (DNN) for their fusion. In order to conceive the temporal quality of human emotion in interactive scenarios, we train in cascade a Long Short-Term Memory (LSTM) layer and a Reinforcement Learning (RL) agent –which monitors the speaker– thus stopping feature extraction and making the final prediction. The comparison of our results on two publicly available A-V emotional datasets viz., RML and BAUM-1s, against other state-of-the-art models, demonstrates the beneficial capabilities of our work.

Where am I? This is one of the most critical questions that any intelligent system should answer to decide whether it navigates to a previously visited area. This problem has long been acknowledged for its challenging nature in simultaneous localization and mapping (SLAM), wherein the robot needs to correctly associate the incoming sensory data to the database allowing consistent map generation. The significant advances in computer vision achieved over the last 20 years, the increased computational power, and the growing demand for long-term exploration contributed to efficiently performing such a complex task with inexpensive perception sensors. In this article, visual loop closure detection, which formulates a solution based solely on appearance input data, is surveyed. We start by briefly introducing place recognition and SLAM concepts in robotics. Then, we describe a loop closure detection system’s structure, covering an extensive collection of topics, including the feature extraction, the environment representation, the decision-making step, and the evaluation process. We conclude by discussing open and new research challenges, particularly concerning the robustness in dynamic environments, the computational complexity, and scalability in long-term operations. The article aims to serve as a tutorial and a position paper for newcomers to visual loop closure detection.

During the last decades, the research endeavours on autonomous driving found great resonance in Advanced Driver-Assistance Solutions that equipped the contemporary civilian vehicles and significantly boosted their driver-less mobility. The existing applications are mostly focused on urban scenarios where signs, road lanes and markers are well defined and ordered favouring the motion of the vehicles whilst, less attention has been paid to the semi-structured and rural environments where traffic infrastructure is scarce. The paper at hand introduces a holistic framework for autonomous vehicles navigation in semi-structured environments. Semantic cues fused with geometrical information of LiDAR data are used for road detection and tracking. OpenStreetMaps are employed as a rough route planner, the waypoints of which are rectified via a probability distribution function over the visible area of vehicle’s vicinity. Thus, vehicle’s localization is obtained by Normal Distribution Transform (NDT) SLAM, where the covariance of egomotion estimation is obtained by processing short-term 3D maps, fused with GPS measurements by means of an extended Kalman filter. Local planning and execution of vehicle’s motion is applied on local cost maps formulated by the union of 2D laser readings and the detected road boundaries fitted through Bézier curves. The complete framework has been evaluated with the aid of a real Autonomous Guided Vehicle in a constrained semi-structured urban area, exhibiting robust navigation performance.

Abstract Simultaneous localization and mapping (SLAM) refers to a process that permits a mobile robot to build up a map of the environment and, at the same time, to use it to compute its location. One of its most important components is its ability to associate the most recently perceived visual measurement to the one derived from previsited locations, a technique widely known as loop closure detection. In this article, we evolve our previous approach, dubbed as ‘DOSeqSLAM’ by presenting a low complexity loop closure detection pipeline wherein the traversed trajectory (map) is represented by sequence‐based locations (submaps). Each of these groups of images, referred to as place, is generated online through a point tracking repeatability check employed on the perceived visual sensory information. When querying the database, the proper candidate place is selected and, through an image‐to‐image search, the appropriate location is chosen. The method is subjected to an extensive evaluation on seven publicly available datasets, revealing a substantial improvement in computational complexity and performance over its predecessors, while performing favourably against other state‐of‐the art solutions. The system’s effectiveness is owed to the reduced number of places, which, compared to the original approach, is at least one order of magnitude less.

Efficient feature learning with Convolutional Neural Networks (CNNs) constitutes an increasingly imperative property since several challenging tasks of computer vision tend to require cascade schemes and modalities fusion. Feature learning aims at CNN models capable of extracting embeddings, exhibiting high discrimination among the different classes, as well as intra-class compactness. In this paper, a novel approach is introduced that has separator, which focuses on an effective hyperplane-based segregation of the classes instead of the common class centers separation scheme. Accordingly, an innovatory separator, namely the Hyperplane-Assisted Softmax separator (HASeparator), is proposed that demonstrates superior discrimination capabilities, as evaluated on popular image classification benchmarks.

Place recognition of pre-visited areas, widely known as Loop Closure Detection (LCD), constitutes one of the most important components in robotic applications, where the robot needs to estimate its pose while navigating through the field (e.g., simultaneous localization and mapping). In this paper, we present a novel approach for LCD based on the assignment of Visual Words (VWs) to particular places of the traversed path. The system operates in real time and does not require any pre-training procedure, such as visual vocabulary construction or descriptor-space dimensionality reduction. A place is defined through a dynamic segmentation of the incoming image stream and is assigned with VWs through the usage of an on-line clustering algorithm. At query time, image descriptors are converted into VWs on the map accumulating votes to the corresponding places. By means of a probability function, the mechanism is capable of identifying a loop closing candidate place. A nearest neighbor voting scheme on the descriptors' space allows the system to select the most appropriate image match at the chosen place. Geometrical and temporal consistency checks are applied on the proposed loop closing pair increasing the system's performance. Evaluation took place on several publicly available and challenging datasets offering high precision and recall scores as compared to other state-of-the-art approaches.

Summary The detection of ambiguous objects, although challenging, is of great importance for any surveillance system and especially for an unmanned aerial vehicle, where the measurements are affected by the great observing distance. Wildfire outbursts and illegal migration are only some of the examples that such a system should distinguish and report to the appropriate authorities. More specifically, Southern European countries commonly suffer from those problems due to the mountainous terrain and thick forests that contain. Unmanned aerial vehicles like the “Hellenic Civil Unmanned Air Vehicle” project have been designed to address high‐altitude detection tasks and patrol the borders and woodlands for any ambiguous activity. In this paper, a moment‐based blob detection approach is proposed that uses the thermal footprint obtained from single infrared images and distinguishes human‐ or fire‐sized and shaped figures. Our method is specifically designed so as to be appropriately integrated into hardware acceleration devices, such as General Purpose Computation on Graphics Processing Units (GPGPUs) and field programmable gate arrays, and takes full advantage of their respective parallelization capabilities succeeding real‐time performances and energy efficiency. The timing evaluation of the proposed hardware accelerated algorithm's adaptations shows an achieved speedup of up to 7 times, as compared to a highly optimized CPU‐only based version.

Summary In this paper, a novel visual Place Recognition approach is evaluated based on a visual vocabulary of the Color and Edge Directivity Descriptor (CEDD) to address the loop closure detection task. Even though CEDD was initially designed so as to globally describe the color and texture information of an input image addressing Image Indexing and Retrieval tasks, its scalability on characterizing single feature points has already been proven. Thus, instead of using CEDD as a global descriptor, we adopt a bottom‐up approach and use its localized version, Local Color And Texture dEscriptor, as an input to a state‐of‐the‐art visual Place Recognition technique based on Visual Word Vectors. Also, we use a parallel execution pipeline based on a previous work of ours using the well established General Purpose Graphics Processing Unit (GPGPU) computing. Our experiments show that the usage of CEDD as a local descriptor produces high accuracy visual Place Recognition results, while the parallelization used allows for a real‐time implementation even in the case of a low‐cost mobile device.

The increasing risks of border intrusions or attacks on sensitive facilities and the growing availability of surveillance cameras lead to extensive research efforts for robust detection of pedestrians using images. However, the surveillance of borders or sensitive facilities poses many challenges including the need to set up many cameras to cover the whole area of interest, the high bandwidth requirements for data streaming and the high‐processing requirements. Driven by day and night capabilities of the thermal sensors and the distinguished thermal signature of humans, the authors propose a novel and robust method for the detection of pedestrians using thermal images. The method is composed of three steps: a detection which is based on a saliency map in conjunction with a contrast‐enhancement technique, a shape description based on discrete Chebyshev moments and a classification step using a support vector machine classifier. The performance of the method is tested using two different thermal datasets and is compared with the conventional maximally stable extremal regions detector. The obtained results prove the robustness and the superiority of the proposed framework in terms of true and false positives rates and computational costs which make it suitable for low‐performance processing platforms and real‐time applications.

In the field of loop closure detection, the most conventional approach is based on the Bag-of-Visual-Words (BoVW) image representation. Although well-established, this model rejects the spatial information regarding the local feature points' layout and performs the associations based only on their similarities. In this paper we propose a novel BoVW-based technique which additionally incorporates the operational environment's structure into the description, treating bunches of visual words with similar optical flow measurements as single similarity votes. The presented experimental results prove that our method offers superior loop closure detection accuracy while still ensuring real-time performance, even in the case of a low power consuming mobile device.

In this paper we propose a novel technique for detecting loop closures on a trajectory by matching sequences of images instead of single instances. We build upon well established techniques for creating a bag of visual words with a tree structure and we introduce a significant novelty by extending these notions to describe the visual information of entire regions using Visual-Word-Vectors. The fact that the proposed approach does not rely on a single image to recognize a site allows for a more robust place recognition, and consequently loop closure detection, while reduces the computational complexity for long trajectory cases. We present evaluation results for multiple publicly available indoor and outdoor datasets using Precision-Recall curves, which reveal that our method outperforms other state of the art algorithms.

Planetary rover exploration missions require accurate and computationally efficient robot localization in order to perform complex and cooperative tasks. The global localization on planetary environments can be competently addressed by incorporating orbital and ground rover imagery. An indicative approach could include (1) the extraction of regions of interest (ROIs) in orbital images, (2) the extraction of ROIs in rover images, (3) the ROI matching, and (4) the localization. In order to perform adequately in ROI matching, a model should be able to detect common ROIs. The work in hand tackles the problem of extracting such regions of interest that are observable on both orbital and rover images. The dedicated model that was designed and implemented contains a detection and a classification part. The detection of the ROIs is based on both their texture and their geometrical properties. Classification was performed on the result of the detection in order to annotate the ROIs and discard any outliers caused by false detection. The results prove that the model is able to detect commonly observable regions and, therefore, is considered to be an adequate preprocessing step in the context of a global rover localization system.

This paper presents a multi-robot system for autonomous vehicle extraction and transportation based on the “a-robot-for-a-wheel” concept. The developed prototype is able to extract vehicles from confined spaces with delicate handling, swiftly and in any direction. The novel lifting robots are capable of omnidirectional movement, thus they can under-ride the desired vehicle and dock to its wheels for a synchronized lifting and extraction. The overall developed system applies reasoning about available trajectory paths, wheel identification, local and undercarriage obstacle detection, in order to fully automate the process. The validity and efficiency of the AVERT robotic system is illustrated via experiments in an indoor parking lot, demonstrating successful autonomous navigation, docking, lifting and transportation of a conventional vehicle.

Space exploratory rovers do well in autonomous or composite semi-autonomous exploration of extraterrestrial surfaces, yet their localization relies on the particular spot they had landed, rather than being universal, i.e. based on the absolute coordinate system of the explored planet. The idea underlaying the work presented in this paper is the transition from the relative to absolute localization by inspecting common Regions of Interest (ROIs) on both rover and orbital imagery. In order to achieve that we propose a method comprising an offline and an onboard procedure. Particularly, prior to the mission the orbital images of the intended landing area are examined to extract ROIs and to construct an offline Global Network (GN). The onboard procedure is based on the rover's self localization which is performed via an inertial aided visual odometry (VO). During its roaming the rover extracts ROIs from the ground and forms a Local Network (LN). The last is iteratively matched with the GN by a specifically designed matching procedure based on Data-Aligned Rigidity-Constrained Exhaustive Search (DARCES). The proposed method is tested on real representative data collected during the ESA Seeker activity. The results indicate that the self-localization of a planetary rover in an absolute frame of reference is feasible, provided that the area includes few discriminative ROIs.

Illegal migration as well as wildfires constitute commonplace situations in southern European countries, where the mountainous terrain and thick forests make the surveillance and location of these incidents a tall task. This territory could benefit from Unmanned Aerial Vehicles (UAVs) equipped with optical and thermal sensors in conjunction with sophisticated image processing and computer vision algorithms, in order to detect suspicious activity or prevent the spreading of a fire. Taking into account that the flight height is about to two kilometers, human and fire detection algorithms are mainly based on blob detection. For both processes thermal imaging is used in order to improve the accuracy of the algorithms, while in the case of human recognition information like movement patterns as well as shadow size and shape are also considered. For fire detection a blob detector is utilized in conjunction with a color based descriptor, applied to thermal and optical images, respectively. Unlike fire, human detection is a more demanding process resulting in a more sophisticated and complex algorithm. The main difficulty of human detection originates from the high flight altitude. In images taken from high altitude where the ground sample distance is not small enough, people appear as small blobs occupying few pixels, leading corresponding research works to be based on blob detectors to detect humans. Their shadows as well as motion detection and object tracking can then be used to determine whether these regions of interest do depict humans. This work follows this motif as well, nevertheless, its main novelty lies in the fact that the human detection process is adapted for high altitude and vertical shooting images in contrast with the majority of other similar works where lower altitudes and different shooting angles are considered. Additionally, in the interest of making our algorithms as fast as possible in order for them to be used in real time during the UAV flights, parallel image processing with the help of a specialized hardware device based on Field Programmable Gate Array (FPGA) is being worked on.

The congregation of crowd undoubtedly constitutes an important risk factor, which may endanger the safety of the gathered people. The solution reported against this significant threat to citizens safety is to consider careful planning and measures. Thereupon, in this paper, we address the crowd evacuation problem by suggesting an innovative technological solution, namely, the use of mobile robot agents. The contribution of the proposed evacuation system is twofold: (i) it proposes an accurate Cellular Automaton simulation model capable of assessing the human behavior during emergency situations and (ii) it takes advantage of the simulation output to provide sufficient information to the mobile robotic guide, which in turn approaches and redirects a group of people towards a less congestive exit at a time. A custom-made mobile robotic platform was accordingly designed and developed. Last, the performance of the proposed robot guided evacuation model has been examined in real-world scenarios exhibiting significant performance improvement during the crucial first response time window.

A new algorithm suitable for the computation of weighted order statistics operations is presented. The algorithm is based on the local histogram and a successive approximation technique. The successive approximation technique ensures that the result of the weighted order statistics operation is computed in a fixed number of steps, equal to the pixel value resolution. Comparative experimental results are also included. It is shown that the proposed algorithm is faster than the quick sort algorithm for weighted order statistics operations.