IET Cybersystems and Robotics最新文献

In the past decades, substantial progress has been made in human action recognition. However, most existing studies and datasets for human action recognition utilise still images or videos as the primary modality. Image-based approaches can be easily impacted by adverse environmental conditions. In this paper, the authors propose combining RGB images and point clouds from LiDAR sensors for human action recognition. A dynamic lateral convolutional network (DLCN) is proposed to fuse features from multi-modalities. The RGB features and the geometric information from the point clouds closely interact with each other in the DLCN, which is complementary in action recognition. The experimental results on the JRDB-Act dataset demonstrate that the proposed DLCN outperforms the state-of-the-art approaches of human action recognition. The authors show the potential of the proposed DLCN in various complex scenarios, which is highly valuable in real-world applications.

Small convolutional neural network (CNN)-based models usually require transferring knowledge from a large model before they are deployed in computationally resource-limited edge devices. Masked image modelling (MIM) methods achieve great success in various visual tasks but remain largely unexplored in knowledge distillation for heterogeneous deep models. The reason is mainly due to the significant discrepancy between the transformer-based large model and the CNN-based small network. In this paper, the authors develop the first heterogeneous self-supervised knowledge distillation (HSKD) based on MIM, which can efficiently transfer knowledge from large transformer models to small CNN-based models in a self-supervised fashion. Our method builds a bridge between transformer-based models and CNNs by training a UNet-style student with sparse convolution, which can effectively mimic the visual representation inferred by a teacher over masked modelling. Our method is a simple yet effective learning paradigm to learn the visual representation and distribution of data from heterogeneous teacher models, which can be pre-trained using advanced self-supervised methods. Extensive experiments show that it adapts well to various models and sizes, consistently achieving state-of-the-art performance in image classification, object detection, and semantic segmentation tasks. For example, in the Imagenet 1K dataset, HSKD improves the accuracy of Resnet-50 (sparse) from 76.98% to 80.01%.

Augmented reality (AR) is gaining traction in the field of computer-assisted treatment (CAT). Head-mounted display (HMD)-based AR in CAT provides dentists with enhanced visualisation by directly overlaying a three-dimensional (3D) model on a real patient during dental treatment. However, conventional AR-based treatments rely on optical markers and trackers, which makes them tedious, expensive, and uncomfortable for dentists. Therefore, a markerless image-to-patient tracking system is necessary to overcome these challenges and enhance system efficiency. This paper proposes a novel feature-based markerless calibration and navigation method for an HMD-based AR visualisation system. The authors address three sub-challenges: firstly, synthetic RGB-D data for anatomical landmark detection is generated to train a deep convolutional neural network (DCNN); secondly, the HMD is automatically calibrated using detected anatomical landmarks, eliminating the need for user input or optical trackers; and thirdly, a multi-iterative closest point (ICP) algorithm is developed for effective 3D-3D real-time navigation. The authors conduct several experiments on a commercially available HMD (HoloLens 2). Finally, the authors compare and evaluate the approach against state-of-the-art methods that employ HoloLens. The proposed method achieves a calibration virtual-to-real re-projection distance of (1.09 ± 0.23) mm and navigation projection errors and accuracies of approximately (0.53 ± 0.19) mm and 93.87%, respectively.

The advancements in intelligent manufacturing have made high-precision trajectory tracking technology crucial for improving the efficiency and safety of in-factory cargo transportation. This study addresses the limitations of current forklift navigation systems in trajectory control accuracy and stability by proposing the Enhanced Stability and Safety Model Predictive Control (ESS-MPC) method. This approach includes a multi-constraint strategy for improved stability and safety. The kinematic model for a single front steering-wheel forklift vehicle is constructed with all known state quantities, including the steering angle, resulting in a more accurate model description and trajectory prediction. To ensure vehicle safety, the spatial safety boundary obtained from the trajectory planning module is established as a hard constraint for ESS-MPC tracking. The optimisation constraints are also updated with the key kinematic and dynamic parameters of the forklift. The ESS-MPC method improved the position and pose accuracy and stability by 57.93%, 37.83%, and 57.51%, respectively, as demonstrated through experimental validation using simulation and real-world environments. This study provides significant support for the development of autonomous navigation systems for industrial forklifts.

A novel correction-enabled Pixel Repetition (PR)-based Reversible Data Hiding (RDH) framework, featuring a new embedding scheme is presented. The proposed RDH scheme uses contextually redundant block pixels, generated via PR, in a two-phase adaptive embedding process, enhancing both image quality and data embedding rates. Specifically, each $����2�\times �2���$ block encodes 4 bits of data using new mapping conditions that facilitate seed pixel reconstruction from remaining block pixels and provide additional embedding opportunities. Additionally, an innovative post-embedding error correction technique, based on $����2^k���$-bit error-correction, minimises post-embedding distortion, further improving image quality. This error correction approach augments data embedding robustness, vital for applications like medical imaging, telemedicine, and digital watermarking that requires high embedding capacity with minimum possible distortion. The proposed scheme surpasses existing state-of-the-art methods in embedding rate-distortion performance, validated through subjective and objective analyses. Furthermore, statistical analysis, including histogram and fragility testing, confirms the scheme's potential for image authentication across diverse multimedia applications. The correction-enabled RDH with PR offers enhanced embedding capacity and image quality preservation, making it particularly advantageous for applications requiring robust data hiding while maintaining visual fidelity.

The phenomenon of sloshing causes a significantly negative impact on a wide range of industries. A time-optimal flatness-based trajectory planning and Lyapunov-based model predictive control (LMPC) is proposed for trajectory tracking of a transmitting cylindrical container filled with liquid. Firstly, this research presents an equivalent discrete model based on a mass-spring-damper system. Subsequently, after the flatness of the adopted non-linear model for 2D is established, time-optimal trajectories are introduced. A control method called LMPC is shown to solve the problem of orbital tracking, which allows setting limits for state variables. In addition, to ensure system performance, a linear extended state observer (LESO) is integrated to cope with system uncertainties. Finally, the efficiency of the proposed approach for liquid sloshing suppression and tracking is illustrated by simulations.

In crowded settings, mobile robots face challenges like target disappearance and occlusion, impacting tracking performance. Despite existing optimisations, tracking in complex environments remains insufficient. To address this issue, the authors propose a tailored visual navigation tracking system for crowded scenes. For target disappearance, an autonomous navigation strategy based on target coordinates, utilising a path memory bank for intelligent search and re-tracking is introduced. This significantly enhances tracking success. To handle target occlusion, the system relies on appearance features extracted by a target detection network and a feature memory bank for enhanced sensitivity. Servo control technology ensures robust target tracking by fully utilising appearance information and motion characteristics, even in occluded scenarios. Comprehensive testing on the OTB100 dataset validates the system's effectiveness in addressing target tracking challenges in diverse crowded environments, affirming algorithm robustness.

Resilient motion planning and control, without prior knowledge of disturbances, are crucial to ensure the safe and robust flight of quadrotors. The development of a motion planning and control architecture for quadrotors, considering both internal and external disturbances (i.e., motor damages and suspended payloads), is addressed. Firstly, the authors introduce the use of exponential functions to formulate trajectory planning. This choice is driven by its ability to predict thrust responses with minimal computational overhead. Additionally, a reachability analysis is incorporated for error dynamics resulting from multiple disturbances. This analysis sits at the interface between the planner and controller, contributing to the generation of more robust and safe spatial–temporal trajectories. Lastly, the authors employ a cascade controller, with the assistance of internal and external loop observers, to further enhance resilience and compensate the disturbances. The authors’ benchmark experiments demonstrate the effectiveness of the proposed strategy in enhancing flight safety, particularly when confronted with motor damages and payload disturbances.

In various fields, knowledge distillation (KD) techniques that combine vision transformers (ViTs) and convolutional neural networks (CNNs) as a hybrid teacher have shown remarkable results in classification. However, in the realm of remote sensing images (RSIs), existing KD research studies are not only scarce but also lack competitiveness. This issue significantly impedes the deployment of the notable advantages of ViTs and CNNs. To tackle this, the authors introduce a novel hybrid-model KD approach named HMKD-Net, which comprises a CNN-ViT ensemble teacher and a CNN student. Contrary to popular opinion, the authors posit that the sparsity in RSI data distribution limits the effectiveness and efficiency of hybrid-model knowledge transfer. As a solution, a simple yet innovative method to handle variances during the KD phase is suggested, leading to substantial enhancements in the effectiveness and efficiency of hybrid knowledge transfer. The authors assessed the performance of HMKD-Net on three RSI datasets. The findings indicate that HMKD-Net significantly outperforms other cutting-edge methods while maintaining a significantly smaller size. Specifically, HMKD-Net exceeds other KD-based methods with a maximum accuracy improvement of 22.8% across various datasets. As ablation experiments indicated, HMKD-Net has cut down on time expenses by about 80% in the KD process. This research study validates that the hybrid-model KD technique can be more effective and efficient if the data distribution sparsity in RSIs is well handled.