Conventional user interfaces (UIs) provide haptic feedback through mechanical components such as buttons, knobs, and rotary components. Haptic feedback remains limited in emerging interactive systems, such as virtual reality (VR), mixed reality (MR), and holographic interfaces. Ultrasonic mid-air haptics (UMAH) enables wearable-free interaction with virtual elements in near-field spaces. However, conventional vibrotactile modulation methods in UMAH systems generate noticeable audible noise and grating lobes that degrade haptic clarity. This work presents a near-field visual-haptic UI system, Holo-haptics, integrating ultrasonic transducer arrays with a dual slit-mirror array (dual SMA) panel to generate a spatially aligned visual-haptic interface. The dual SMA panel provides floating visual UI elements and a reflected ultrasonic path, enabling multi-directional haptic feedback around the user’s finger. Three driving patterns are designed to emulate the tactile sensations of traditional UI components, including buttons, knobs, and sliders. Triangular pulse-width modulation (TPWM) and a directivity-based driving mask are implemented to mitigate audible noise, suppress grating lobes, and lower power consumption. Experimental results show that the driving mask reduces the grating-lobe intensity by 51.82% while maintaining 92.85% of the focal-point pressure. Under the same perceived feedback intensity, TPWM reduces driving power consumption to 88.33% of that required by spatiotemporal modulation (STM). These results support the use of Holo-haptics for near-field wearable-free visual-haptic UI interaction.
Rapid population aging presents significant societal challenges, including a high risk of domestic accidents among elderly individuals, particularly in bathrooms. These incidents, such as falls and medical emergencies, can have severe consequences if not detected promptly. Although monitoring solutions such as RGB(-D) cameras and LiDAR are available, they are often unsuitable for home environments owing to their high cost and intrusive nature. As a low-cost alternative, this proof-of-concept study investigates bathroom-related behavior estimation using sound event detection (SED) based on environmental audio captured by an inexpensive microphone array. We developed a dataset comprising seven bathroom-related behavioral events, including falls, showering, and impacts. We propose a Conformer architecture with exponential moving average (EMA) teacher-student consistency regularization, trained on log-mel spectrogram features to classify these acoustic events. To reduce false alarms and prediction instability, we employ frame aggregation for 1 s temporal smoothing. The system was evaluated under three conditions: isolated-event, matched-conditions sequential, and mismatched-conditions sequential tests conducted in two physically distinct bathrooms. With frame aggregation, the model achieved macro-average F1-scores of 93.5 ± 1.9%, 79.9 ± 2.4%, and 78.4 ± 5.3% on the respective tests. These scores were higher than those of a frozen AudioSet-pretrained AST baseline at the same resolution under mismatched conditions. Although further validation with larger and more diverse datasets is required, the results demonstrate the feasibility of Conformer-based SED with post-processing for low-cost bathroom activity monitoring.
Low Altitude Economy (LAE) is rapidly transforming consumer-facing and mission-critical aerial services enabled by agentic AI in order to deliver adaptive analytics and secure, trustworthy edge intelligence. In this paper, we introduce an agentic edge-intelligence framework for emergency drone safe landing in dynamic LAE environments by jointly optimizing (i) the drone’s safe-landing mobility decisions and (ii) an Integrated Sensing and Communication (ISAC) power split, while explicitly accounting for the drone’s energy availability, communication reliability, and secrecy requirements. The environment is modeled as partially observable by the drone, which maintains a belief map over the safety of the terrain in terms of landing. A physics-aware emergency energy model is introduced for the drone to capture its energy consumption during its hover, drag, turning, and descent-related components, as well as the dynamics of the environment. We formulate the resulting decision process as a Constrained Markov Decision Process (CMDP) with multi-dimensional costs capturing crash events, unsafe-terrain exposure, SINR outage, secrecy deficits against eavesdropping, and cumulative energy usage. The CMDP is solved via a primal–dual Lagrangian PPO approach that updates both the policy and Lagrange multipliers to enforce constraint satisfaction online under highly dynamic LAE conditions. A detailed evaluation of the proposed framework is performed, showing that the proposed agentic ISAC-aware emergency safe-landing framework achieves superior landing success with low violation rates and bounded per-step energy consumption, and adaptively adjusts its ISAC power split to sustain secrecy throughout descent, compared to existing approaches.
With the rapid development of 3D imaging and display technologies, point clouds have become a fundamental data representation in a wide range of applications, including autonomous driving, virtual reality, etc. Nevertheless, during the processes of acquisition, compression, transmission, and reconstruction, point clouds are unavoidably subject to various distortions, which may result in significant degradation of perceptual quality. To ensure both a satisfactory user experience and reliable performance in downstream tasks, Point Cloud Quality Assessment (PCQA) is becoming increasingly critical. Most PCQA methods have largely focused on the content-aware factor using 2D projections or 3D descriptors, with insufficient attention given to the distortion-aware factor. To confront this, we propose a novel distortion-aware quality perceptual modeling framework for No-Reference PCQA (NR-PCQA), mainly by decoupling multi-grained distortions and then interacting these decoupled distortion-aware features to regress quality labels. Concretely, the proposed framework comprises a distortion-degree modeling branch trained by supervised contrastive learning and a distortion-classification modeling branch implemented through a classification network, which generate distinct features, similar to a divide-and-conquer strategy, representing vague quality judgment and degraded scene perception, respectively. Following this, a structure- and degree-guided feature correction network is proposed to fuse layered distortion features. It utilizes geometric cues as attention weights for feature refinement and employs a gating mechanism to enable multi-grained feature interaction. Finally, these features are mapped to distortion-degree and distortion-type correlated scores via regression respectively, which are then summed to yield the final score. Extensive experiments across three datasets (WPC, SJTU and LS-PCQA) demonstrate that our method significantly outperforms state-of-the-art approaches across most evaluation metrics. The ablation studies further validate the rationality of our dual-pathway interaction framework. The code will be available at https://github.com/wanghan-nbu/Multi-Grained-Distortion-Decoupled-PCQA.
Ensuring reliable security monitoring in modern consumer electronic systems requires accurate video anomaly detection models capable of learning complex spatial and temporal behavioral patterns. To address the limitations of existing weakly supervised methods in capturing segment-level saliency and temporal continuity, this paper proposes a Dual-Branch Feature Fusion-Based Constrained Multi-Loss Video Anomaly Detection framework, termed DBF-CML-transMIL. The proposed framework integrates multiple instance learning with a dualbranch architecture, where a multi-layer linear network extracts discriminative spatial saliency features from video segments, while a cascaded Transformer-based attention module captures multi-level temporal dependencies and contextual correlations among segments. In addition, a constrained multi-loss optimization strategy combining sliding-window MIL loss and center loss is introduced to improve feature discrimination, robustness, and prediction diversity. To overcome the limitations of conventional discrete top-k optimization, a constraint-guided sliding-window top-k mechanism is further designed to preserve temporal continuity and reduce noisy instance selection during training. Extensive experiments and ablation studies conducted on the UCSD Pedestrian and UCF-Crime datasets demonstrate the effectiveness of the proposed framework, achieving AUC scores of 97.56% and 84.26%, respectively. Experimental results confirm that the proposed dual-branch learning and constrained optimization strategies significantly improve anomaly representation capability and detection performance, making the framework suitable for intelligent consumer electronic surveillance applications requiring real-time and reliable anomaly detection.
