The increasing use of robot applications in various industries requires close monitoring and management of the data generated by these systems. Therefore, it is essential to implement a monitoring system that identifies and reports abnormal situations in robots. The literature differentiates between rule-based and machine learning methods. Rule-based approaches rely on predefined rules to detect deviations from expected behavior. In contrast, machine learning algorithms acquire the ability to learn patterns on their own. This paper evaluates both approaches to determine whether machine learning algorithms can replace or enhance rule-based methods. The evaluation employs an actual production scenario that deals with three induced anomalies: Additional weight, drop of the manipulated object, and reduced speed.
Material failure is frequently caused by material corrosion. To expedite maintenance by enabling more accurate identification of corroded areas in materials, this study employs a corrosion target detection image dataset, using 12 types of image enhancement and image filtering methods to preprocess the image dataset and using the You Only Look Once (Yolo)v8n model for image detection on each preprocessing dataset. The simulation's findings demonstrate that in contrast to Yolov5n and Yolov6n, Yolov8n can achieve higher Precision, Recall, mAP50, and mAP50-95 values. Simultaneously, maximum Recall and mAP50 values were obtained using color enhancement with a multiplier of 1.5, maximum mAP50-95 value was obtained using median filtering with a convolutional kernel size of 3x3, and maximum Precision value was obtained using minimum filtering with a convolutional kernel size of 5x5. The above results illustrate that selecting appropriate image preprocessing methods can further improve the yolov8 model for identifying corrosion areas in the image. Thus, these findings can provide a certain reference value for the field of corrosion detection.
This paper develops a three-degree-of-freedom rehabilitation robot and establishes its mathematical model along with a Permanent Magnet Synchronous Motor (PMSM) servo control system. Using Simulink software, the current-loop control is implemented to achieve torque control and perform simulation analysis. Additionally, to overcome the effects of gravity, friction, and inertia on the motor during rehabilitation, compliance control is edited via the teach pendant to ensure smooth operation. Motor data, including rotational angle, angular speed, and instantaneous current changes, are captured during the process. Finally, the measured data from the robot's drive system are compared with the Simulink simulation results to verify the effectiveness of the proposed approach.
In order to solve the problem of low accuracy of traditional UAV detection due to RF failure, electromagnetic interference, complex scenes, etc., this paper proposes a UAV detection method based on transfer learning and YOLOv5-Lite model, aiming to carry out efficient and accurate multi-target detection through computer vision technology. In this paper, a diverse dataset containing flying targets such as drones and birds is constructed, and the pre-trained model is optimized and trained on the self-built dataset enriched by data augmentation technology, so that it can cope with UAV detection tasks in various complex natural scenes. Experimental results show that YOLOv5-Lite has excellent performance in complex environments, especially in multi-target and small-target detection tasks, with high accuracy, low false alarm rate and fast inference ability. Compared with similar models, the proposed method has significant advantages in terms of detection accuracy, inference speed and model parameter quantity, and is especially suitable for scenarios with high real-time and resource-constrained applications.
In modern unmanned vehicle technology, outdoor patrolling algorithms play a crucial role in enabling autonomous navigation. To enhance the performance of unmanned vehicles in complex outdoor environments, researchers have been continuously refining and optimizing these algorithms. This paper presents a detailed overview of our approach, including the design of the autonomous navigation camera algorithm, vision processing, and recognition technologies. To address challenges such as limited computational resources, environmental interference in line patrolling, and the prolonged debugging time required for the line patrolling algorithm, we propose innovative solutions. These include multi-threading, segmented adaptive binarization, the edge inheritance method, the least squares approach, and enhancements to the YOLOv5-Lite model. Together, these advancements significantly improve the vehicle's adaptability to diverse lane conditions.
This paper investigates a path planning approach for cooperative unmanned aerial vehicle (UAV) and unmanned ground vehicle (UGV) systems. Considering the position error corrections (PEC) of UAV, threats, and other complex constraints, a UAV-UGV cooperative path planning model is developed to minimize the total cost. Then, a Hybrid Ant-Pigeon Cooperative Optimization (HAPCO) algorithm is presented to address issues in existing bio-inspired optimization algorithms. A new pheromone update strategy and adaptive inertia weight are designed to optimise the performance. the performance of the Ant Colony Optimization (ACO) algorithm and Pigeon-inspired Optimization (PIO) algorithm. Moreover, the two algorithms are integrated to perform a full-cycle search. Simulation results reveal that the HAPCO algorithm outperforms ACO and PIO regarding path quality, cooperation cost, convergence, and computational efficiency, further demonstrated the validity and superiority of our algorithm.
Plant diseases pose a significant threat to global food security, with annual losses amounting to billions of dollars. Early detection and intervention are critical for mitigating these losses. Traditional methods of disease detection, such as visual inspection, are time-consuming, labor-intensive, and often inefficient. This paper presents an advanced approach using the YOLOv5 (You Only Look Once) model, a state-of-the-art object detection algorithm, to identify and classify plant diseases in real-time. The YOLOv5 model offers high accuracy and efficiency, particularly for tasks requiring simultaneous detection of multiple objects, making it ideal for agricultural applications. This paper explores using novel data collection methods, such as mounting cameras on Husky UGV (Unmanned Ground Vehicle), to capture diverse images of crops from various angles in the field. These autonomous systems enhance dataset variety, contributing to more accurate detection results. Our methodology focuses on training the YOLOv5 model using annotated datasets, which enables the model to identify diseases such as apple rust leaf, bell pepper leaf spot, and others. The experimental results demonstrate a detection accuracy of 92 %, showcasing the potential of YOLOv5 for realtime crop disease monitoring. Integrating robotics for field analysis is further discussed as a future direction, providing a scalable solution for continuous crop monitoring and disease management.
Cross-modal retrieval has always been one of the essential tasks in multimodal learning, with the aim of searching for the most relevant cross-modal samples based on a given query. Taking image-text retrieval as an example, a direct and efficient approach is to use visual semantic embedding technology to bring positive samples closer and push away negative samples in the shared subspace. However, these methods often overlook the potential implicit alignment relationships in text. In this letter, we enhance representation by revisiting potential alignment relationships in text and propose a new method called IACR that exploits an Implicit Alignment Module (IAM), thereby capturing the fine-grained local associations. Specifically, IAM captures implicit alignment through a multi-head attention mechanism to establish fine-grained alignment between image representations and partial words in the text. This approach avoids the additional reasoning costs caused by fine-grained modeling and maintains local correlation to improve performance. To verify the superiority of this cross-modal alignment paradigm, we conduct experiments on two large-scale image-text retrieval datasets, i.e., Flickr30K and MS-COCO, to evaluate the proposed IACR and provide analysis to verify its effectiveness.
The fusion of digital and physical worlds through augmented reality (AR) has transformed human interaction with information and the environment, all while prioritizing design for humanity to ensure accessibility, inclusivity, and well-being. This study investigated an AR-incorporated machine learning application named “matngon” aimed at detecting and mitigating everyday risks such as stairs, windows, scissors, knife, outlet and doors. Anchored on the Protection Motivation Theory and the Technology Acceptance Model, the study aimed to address the unavailability and lack of accessibility and usability of AR applications (apps) for risk detection, particularly countries like the Philippines. The system integrates object detection algorithms using Tensorflow.js with real-time camera feeds on mobile devices to provide proactive hazard identification and warnings. Additionally, it provides features like first aid modules and an emergency contacts list to make it an integrated risk management application. The performance and functionality testing of the app demonstrates efficient execution of various features and successful completion of risk detection which is also reflected in the users' positive perceptions regarding the application's usefulness, ease of use, and intention to adopt based on the user testing and interviews. However, the considerations of the following are recommended for future research: enhancing detection accuracy by implementing transfer learning for model refinement, expanding detection classes, and improving user interface and additional features to enhance user experience and accessibility. Thus, matngon demonstrates promising potential as a proactive risk detection and management tool, with scope for further refinement and enhancement to ensure usability and effectiveness in promoting safety and well-being.
Semantic segmentation on point cloud plays an important role in robotic applications. However, current point cloud interpolation methods have shortcomings. Distance-based interpolation schemes may confuse different semantic information, while feature-based interpolation schemes suffer from unstable training processes. To tackle this issue, we propose Category-Augmented Net (CAN). The key module of CAN is Category-Augmented Interpolation (CAI), which is used for point cloud upsampling. CAI applies Category-Augmented KNN to find the point set for interpolation in a hierarchical manner, which takes both distance and semantic information into account. Furthermore, we also introduce a dynamic training strategy, allowing the structure of our model to change during training, which makes the training progress much faster. Specifically, in the early stage of training, CAI relies on ground truth categories to select interpolation point sets, while in the later stage, CAI utilizes the predicted feature vector for this selection process. Extensive experiments on Semantic Cluttered Objects (SCO) show the effectiveness of our method.
Perceptual coding algorithms play a crucial role in eliminating visual redundancy in video compression. Despite their promising results, there remains potential for further optimization, particularly in leveraging human visual characteristics for rate control and effectively applying just noticeable distortion (JND)a threshold defining human visual sensitivity to changes. In this paper, we propose a novel HEVC rate control scheme inspired by the focusing mechanism of the human visual system (HVS). By integrating saliency and focal distance information, we construct a focusing factor and incorporate it into the encoding framework. This focusing factor serves as a key parameter for rate control, optimizing LCU-level bit allocation and enhancing video coding efficiency. Furthermore, we combine the focusing factor with a JND-based model to design a Gaussian filter that smooths nonfocused areas, reducing background redundancy while enhancing visual quality. The proposed algorithm is implemented on HM16.7, and experimental results show bitrate savings and PSNR improvements ranging from 0.1 to 0.5 dB compared to state-of-the-art algorithms. Additionally, subjective evaluations confirm significantly enhanced visual quality at equivalent bitrates.
Rib fractures are common injuries that pose significant risks, necessitating rapid and accurate diagnosis. Traditional chest CT radiographic diagnostic techniques are highly dependent on the manual reading of images. The process is labor-intensive, highly experience-dependent, and less accurate, prompting the need for advanced methods based on artificial intelligence. To solve this problem, this study demonstrates a novel automated rib multiplanar reconstruction method that integrates object detection and tracking algorithms, specifically YOLO-X and ByteTrack. A dataset of 200 patients from Tongji Medical College, including 88 with rib fractures and 112 normal, is utilized to train and validate the algorithm model. Superior performance with high precision (0.993), recall (0.995), and mAP (0.994) can be achieved by YOLOv10-ByteTrack integration. The tracking algorithm improves the success rate by over 30 %, with a statistical error reduction of nearly 70%, greatly enhancing the detection accuracy. This study represents a significant step forward in computerized tomography reconstruction, provides a reliable solution for rib fracture diagnosis, and addresses limitations of traditional MPR, contributing to both scientific research and clinical practice.
Addressing uncertainty is a challenge in decisionmaking. Soft set theory aids in complex scenarios involving multiple opinions. We propose a game-theoretic soft set model to resolve conflicts in soft sets, particularly when merging multiple classifiers. This model allows each classifier to make strategic selections from objects associated with conflicting opinions. The experimental results demonstrated that the model effectively balances the decisions of all classifiers while maintaining satisfactory classification performance.
Road cavities, often called potholes, pose a major problem for our transportation infrastructure, causing accidents, costly repairs, and traffic jamming. This study shows using a computer vision technique called You Only Look Once (YOLO) to find potholes on roads. YOLO is proficient and efficient for realtime object detection. The researchers want to see which version of YOLO is more effective in object analysis in pictures and videos. We'll be comparing YOLOv5, YOLOv7, and YOLOv8, thereby also setting up the system in Gazebo simulated environment. By carefully analyzing the results, we aim to understand the advantages and disadvantages of each YOLO version. This will help us see the most useful ones for finding potholes when managing and maintaining roads.
Recent advances in contrastive learning promote the research of many downstream tasks, especially deep clustering, which explores the potential semantic connections for unlabeled samples. However, these contrastive-based clustering methods focus on positive pairs in the pairwise contrastive framework and ignore the latent semantic relations of negative pairs, causing semantic information distortion in embedding space. In this paper, we propose joint distribution contrastive learning (JCL), an unsupervised image clustering method encoding semantic structures of negative pairs into the learned embedding space. Specifically, JCL introduces latent class variables and model discrimination task as a maximum average class conditional likelihood estimation to encourage negative pairs with the same semantic information to be closer in embedding space. The proposed joint contrastive loss of JCL is the negative-wise contrastive loss and serves as the objective function of deep clustering. JCL is a simple end-to-end online deep contrastive clustering method that jointly exploits the positive and negative pairs and synchronously learns representation and clustering to optimize the network. Extensive experiments on moderate-scale image clustering benchmarks demonstrate JCL remarkably outperforms the state-of-the-art methods.
We present a full-stack autonomous navigation system for a reconfigurable tracked robot in man-built environments, especially targeting the safe traversing in stairways. This system first uses A* algorithm to plan an initial path across the floor on a grid map. Then the staircase detection algorithm will continuously look for the staircase features from the environment. Once the staircase is detected, our motion strategy completes the staircase traversing, otherwise a follower based on dynamic window approach is used to follow this path. Finally, the low-level controller drives the motor to rotate to complete the cross-floor navigation task. Our method has been validated experimentally on a reconfigurable tracked robot. Compared to manual control by experienced professional operators, our approach enables the robot to complete the cross-floor traversing task more stably, smoothly and safely. Overall, this work combines the advantages of different algorithms in cross-floor navigation, improves the stability and adaptability of robots in different environments and provides important experience for future research.
In response to the problem of the filtering error of the CKF algorithm increasing linearly with the dimensionality of the state space, resulting in difficulty in propagating multiplicative noise, and the instability of the system caused by outliers in the system state driven model and observation data, this paper proposes an Extended Dimensional Embedded Cubature Kalman Filter algorithm based on the Truncated Singular Value Decomposition (TSVD-AECKF). Firstly, Singular Value Decomposition (SVD) is used instead of Cholesky Decomposition in the CKF algorithm to suppress the non-positivity of the system state covariance matrix; Considering the impact of small singular values on stability, this paper adopts the truncation method and provides a method for determining the truncation threshold; Secondly, the system noise is added to the state variable, and the embedded cubature criterion is used to improve the traditional CKF while expanding its dimensions; Finally, through simulation experiments, TSVD-AECKF was compared with other SLAM methods, and the results showed that this method can effectively suppress positioning errors that increase with the dimension of the state space, enhance the filter's ability to resist noise data interference, and thereby improve the robustness and stability of the mapping.
Convolutional neural networks have shown excellent performance in image processing. However, due to the black-box property of deep networks in applications, people still doubt its credibility, which has led to the birth of many interpretation methods for computer vision networks, but their interpretation effects can only be subjectively assessed by human vision on the heatmaps generated by these interpretation methods. In this paper, we propose a novel Geospatial Analysis for eXplainable Artificial Intelligence (GeoSeXAI) method, which for the first time introduces the spatial autocorrelation analysis method in geostatistics combined with the XAI algorithm and the SAR images to provide an accurate and objective quantitative assessment of the spatial distribution of the XAI attribution, as well as a global and objective assessment of the XAI attribution. global objective quantitative assessment, and provides a local qualitative explanation that is easier for human visual understanding.
The live detection system for tensioning clamps based on unmanned aerial vehicles is the development direction of routine inspections for high-voltage transmission lines. The detection of tensioning clamps in real-time in complex environments is the fundament of this system. Addressing this problem, YOLOv8-SC is proposed based on YOLOv8. Replacing the original C2f module with the new C2fG-Ghost module, and a GAM attention layer is added to the backbone network. A binocular 3D coordinate detection algorithm is added to obtain the relative position of the target. Experiments show that the improved algorithm improves by 2.7% and 7.0% on mAP50 and mAP50-95, respectively, and the F1-score improves by 2.4%. Experiments on-site verified that this algorithm could loaded on a portable intelligent terminal with a detection speed of 89 FPS, which satisfied the real-time demands.
Recently, research on the flight control of indoor small unmanned airships has attracted much attention. Unmanned airship has characteristics of low manufacturing cost, long dwell time, low energy consumption and so on. Under the premise of ignoring the external wind interference, a controller is designed that makes the airship move to the preset point and stay in the air. The mathematical model of the unmanned airship is first introduced in this paper. A six-degree-of-freedom equation of motion is established after force analysis, and then it is simplified to a three-degree-of-freedom model according to the geometric symmetry of the airship. Subsequently, the motion process of the airship is disassembled into horizontal motion and vertical motion, the backstepping method is used to design the controller, and unknown parameters in the model are processed with adaptive control. Finally, computer simulation experiments are carried out to verify the correctness of the model and the effectiveness of the controller from a mathematical point of view. In this paper, a small unmanned airship motion control system is designed, and its feasibility from the perspective of computer simulation is verified. These studies provide initial exploratory work on control methods for indoor unmanned airships.
