Yangyang Wang

Human pose estimation with millimeter-wave radar has become a promising alternative sensing method due to its robustness against varying lighting conditions and occlusions. However, existing methods often suffer from temporal inconsistencies and spatial inaccuracies, primarily due to the sparse nature of point clouds and the underutilization of valuable Doppler velocity information. To address this challenge, we propose VCF-Pose, a novel Velocity-guided Coarse-to-Fine Pose estimation framework that creatively transforms unstructured Doppler data into a structured motion representation to dynamically guide the refinement of an initial coarse pose, ensuring it aligns with temporal dynamics and anatomical priors. Specifically, VCF-Pose first generates an initial coarse pose estimate using a GRU-based sequential model to capture temporal context from voxelized radar data. Then we design a velocity space projection module that creatively transforms unstructured raw Doppler data into a structured 3D velocity feature volume. This novel representation enables a 3D relational convolution network to effectively model the fine-grained correlations between motion dynamics and spatial geometry. Finally, a structural prior-guided feature fusion module leverages anatomical constraints to perform global alignment and refinement, ensuring the final pose is anatomically plausible. Extensive experiments demonstrate that VCF-Pose significantly improves inter-frame continuity and reduces estimation error, achieving an average error of 4.03 cm and outperforming existing state-of-the-art methods.

Reliable trajectory estimation and precise motion control are the prerequisites for underwater robotic systems to perform complex autonomous tasks, which are essential for enhancing the operational efficiency of intelligent underwater facilities. However, the inherent asymmetry of underwater hydrodynamics, featureless images caused by complex environments, and the lack of high-frequency state feedback significantly hinder stable trajectory tracking and robust autonomous navigation. To address these challenges, this paper proposes an integrated autonomous navigation and robust control scheme for underwater robotic systems. Specifically, we first propose a visual–inertial trajectory estimation method for underwater robotic systems, which effectively overcomes the challenges of featureless images and provides consistent, real-time pose feedback for motion execution. Furthermore, we develop a hierarchical robust motion control strategy for autonomous underwater robots, which integrates model predictive control with incremental nonlinear dynamic inversion to achieve precise positioning performance and reliable operation under environmental disturbances. Finally, we design and implement a customized, highly integrated underwater robotic platform that integrates the proposed trajectory estimation and robust control modules, with its performance validated through extensive field experiments in underwater scenarios. The experimental results demonstrate that the proposed system can effectively achieve high-precision trajectory tracking and maintain operational stability, providing a comprehensive engineering solution for the autonomous navigation of underwater robots in complex environments.

Environmental perception is a cornerstone for autonomous underwater vehicles (AUVs) to achieve robust self-localization and scene understanding, which are pivotal for the intelligent management of marine ranching. However, underwater image degradation and weak-textured scenes significantly hinder reliable self-localization and fine-grained environmental perception. To address the perceptual asymmetry arising from these challenges, this paper proposes a robust visual–inertial simultaneous localization and mapping (SLAM) and biomass assessment scheme for marine ranching. Specifically, we first propose a robust tightly coupled underwater visual–inertial localization scheme, which leverages a multi-sensor fusion strategy to solve the image degradation problem of localization in complex underwater environments. Furthermore, we propose a novel underwater scene perception method, which enables the simultaneous visual reconstruction of aquaculture species and the quantitative mapping of their spatial distribution in marine ranching. Finally, we develop a low-cost, agile, and portable multisensor-integrated system that consolidates autonomous localization and aquaculture biomass assessment modules, with its performance validated through extensive real-world underwater experiments. The experimental results demonstrate that the proposed methods can effectively overcome the interference of complex underwater environments and provide high-precision perception support for both AUV state estimation and aquaculture asset management.

In Internet of Things (IoT) applications, privacy-preserving concerns restrict direct data sharing across distributed devices. Moreover, disparities in device performance, operational environments, and sensing capabilities lead to highly imbalanced data distributions across clients, undermining model generalization and limiting the adaptability to diverse data distributions. While federated learning (FL) enables collaborative training without raw data exchange, it remains vulnerable to imbalanced data distributions across clients, often resulting in global model bias and poor generalization. Existing methods mitigate this by sharing partial statistical representations, yet fail to comprehensively capture complex data distributions. To address this issue, we propose uncertainty-guided Gaussian representation sharing (UGRS), a novel FL method designed to capture more comprehensive distributional characteristics. UGRS shares global Gaussian representations to guide local training on clients and dynamically refines these representations based on sample uncertainty. Gaussian representations inherently encode holistic data distribution features, making them more effective than partial statistical features in reflecting the complexity of the entire data distribution. In addition, UGRS constructs compact Gaussian representations by projecting high-dimensional features into a low-dimensional space—this not only reduces data transmission overhead but also minimizes privacy leakage risks by limiting exposure of sensitive high-dimensional information. Experiments on MNIST, CIFAR-10, CIFAR-100, Tiny-ImageNet, and a custom IoT dataset with client-specific imbalanced distributions demonstrate that the UGRS consistently outperforms state-of-the-art methods with an average accuracy improvement of 0.87%–5.72% for both global and local models.

Ensuring resilient semantic segmentation under diverse outdoor conditions is vital for autonomous driving. However, nighttime segmentation remains underdeveloped compared to its daytime counterpart due to poor illumination and scarce annotated data, posing a significant challenge for night-scene understanding. Most current approaches mainly rely on domain adaptation technologies to transfer segmentation models trained on daytime scenes. However, the substantial distinctions between daytime and nighttime domains often hinder effective adaptation. To address this challenge, we leverage the implicit comprehensive information within nighttime data to enhance semantic segmentation through a semi-supervised approach. Specifically, we introduce a Semi-Supervised Comprehensive Learning (SSCL) approach, which is a unified, closed-loop learning architecture composed of three mutually reinforcing correction mechanisms: (1) unsupervised interactive correction to mitigate the risk of erroneous label propagation by leveraging complementary learning abilities; (2) unsupervised reinforcement correction, which enhances the model’s adaptability by promoting diverse learning on high-uncertainty regions through entropy-guided perturbation; (3) supervised standard correction to ensure alignment with known standards by anchoring the system to reference answers. SSCL is the first semi-supervised algorithm that jointly exploits structural diversity, uncertainty-aware supervision, and closed-loop correction to fully harness the latent potential of unlabeled nighttime data. Extensive experiments on NightCity, Dark Zurich and Nighttime Driving datasets demonstrate that SSCL achieves state-of-the-art performance in nighttime semantic segmentation.

Visual trajectory estimation can endow autonomous underwater vehicles (AUVs) with environmental perception capabilities and has broad application prospects in the fields such as oceanographic surveys, underwater construction, and marine ranching. However, due to featureless images and depth ambiguity issues caused by underwater multiple mediums environments, monocular visual trajectory estimation in complex underwater environments remains a challenging problem. In this paper, we propose a monocular visual trajectory estimation method for AUVs, which can address the challenges of featureless and depth ambiguity by leveraging deep image representations and multi-view geometry. Specifically, we design a bidirectional optical flow consistency scheme that selects sparse correspondences from monocular dense predictions to deal with the featureless images, and then achieve AUV trajectory estimation through epipolar constraints. Furthermore, we propose an iterative depth-consistent method, which solves the problem of depth ambiguity by aligning geometrically triangulated depths to the scale-consistent deep depths. We also develop a low-cost, agile, and portable AUV picking system with real-time trajectory estimation capabilities, and carry out extensive experiments in the Yellow Sea to test its performance. The experimental results demonstrate the effectiveness of the proposed method.

Neural implicit visual simultaneous localization and mapping (SLAM) systems demonstrate high-fidelity mapping and robust tracking performance in static environments, making them a crucial solution for real-time dense reconstruction. However, in the real world, dynamic objects often interfere with the map construction and localization accuracy of the SLAM system, resulting in significant drift and artifacts. To address these issues, this paper combines the strengths of traditional semantic SLAM and neural implicit SLAM by utilizing semantic segmentation and Conditional Random Fields (CRF) to detect and segment dynamic and static objects, preserving static points while evaluating the dynamic probability of pixels. We propose a keyframe selection and background inpainting strategy that significantly reduces reconstruction artifacts and enhances tracking robustness. Additionally, we introduce a depth penalty based on pixel dynamic probability to control rendering loss and mitigate the impact of transient interference. Our method was evaluated on the TUM RGB-D and Bonn datasets with dynamic interference and compared against existing traditional dynamic semantic SLAM and neural implicit SLAM systems. Experimental results show that our SLAM system maintains robust tracking and produces high-quality reconstructions, achieving a 90% reduction in average trajectory error (ATE) compared to state-of-the-art neural implicit SLAM systems.

Underwater image enhancement plays a pivotal role in addressing the challenges posed by the complex and dynamic underwater environment. While previous research has conducted valuable explorations from a global enhancement perspective, underwater settings often exhibit multi-distribution characteristics in both spatial frequency and illumination conditions that require specialized attention, that is, multiple spatial frequencies and lighting conditions coexist in the same image, making it difficult to achieve optimal enhancement using global mapping. To address this challenge, we propose a Multi-Distribution Aware Network (MDA-Net) that leverages local frequencies and illumination characteristics of images for adaptive adjustment to balance the diverse visual enhancement requirements of local regions. Specifically, to address the challenge of multiple spatial frequency distributions, we explore the correlation between spatial frequency, receptive field, and image quality perception, and design a Frequency-Aware Kernel Selection Convolution, which could adaptively select the size of convolution kernels based on the frequency complexity of each region, so as to balance the requirements of noise reduction and color fidelity in different regions. Furthermore, to address the challenge of multiple illumination distributions, we leverage the inherent illumination characteristics of the image to generate a Gamma Transformation-based Illumination Balancer, whose neurons can comprehensively perceive global and local illumination through multi-parameter correction representation, thereby guiding the focus of the enhancement work. Extensive experiments with ablation analysis show the effectiveness of our proposed MDA-Net on four benchmark datasets: UFO-120, UIEB, UIEB-U60 and U45.

Camera and IMU are widely used in robotics to achieve accurate and robust pose estimation. However, this fusion relies heavily on sufficient visual feature observations and precise inertial state variables. This paper proposes PLE-SLAM, a real-time visual-inertial SLAM for complex environments, which introduces line features to point-based SLAM and propose an efficient IMU initialization method. Firstly, we use parallel computing methods to extract point-line features and compute descriptors to ensure real-time performance. Adjacent short-line segments are merged into long-line segments for more stable tracking, and isolated short-line segments are directly eliminated. Secondly, to overcome rapid rotation and low texture scenes, we estimate gyroscope bias by tightly coupling rotation pre-integration and 2D point-line observations without 3D point cloud and vision-only rotation estimation. Accelerometer bias and gravity direction are solved by an analytical method, which is more efficient than non-linear optimization. To improve the system’s robustness in complex environments, an improved method of dynamic feature elimination and a solution for loop detection and loop frames pose estimation using CNN and GNN are integrated into the system. The experiment results on public datasets demonstrate that PLE-SLAM achieves more than 20%~50% improvement in localization performance than ORB-SLAM3 and outperforms other state-of-the-art visual-inertial SLAM systems in most environments.