The training of end-to-end models for audio-driven 3D facial animation is often hindered by the significant computational overhead of direct loss computation on high-dimensional 3D mesh data. To address this efficiency bottleneck, we propose a novel compressed-domain training framework. Our method leverages the principles of Compressive Sensing to shift the loss computation from the high-dimensional vertex space to a compact, low-dimensional signal space. Extensive experiments on multiple state-of-the-art architectures demonstrate that our framework significantly accelerates the training process, with speedups ranging from 1.52x to a remarkable 9.57x, while maintaining generation quality comparable to SOTA methods. Our findings establish a superior solution that achieves an exceptional trade-off between computational efficiency and final animation quality.

This work investigates Integrated Communication and Sensing (ICAS) for intelligent transportation systems (ITS), with a focus on pedestrian and vehicle detection to enhance road safety in high-mobility scenarios. We review system and sensing channel models alongside state-of-the-art monostatic and bistatic sensing algorithms, analyzing their strengths and limitations. The Iterative Strong Target Cancellation (ISTC) algorithm is examined as an effective method to mitigate interference and enable reliable multiple-target detection. Building on this foundation, we propose a novel sensing-aided trajectory prediction algorithm that supports proactive beamforming, significantly improving communication reliability. Simulation results demonstrate that predictive and interference-resilient sensing techniques allow ICAS to meet stringent ITS requirements. A high-fidelity 3D simulation framework and practical algorithmic insights are also provided, positioning ICAS as a key enabler for future 6G networks.

Multiple-object tracking (MOT) involves identifying and consistently assigning IDs to multiple targets in a video sequence, where occlusions are frequently encountered. Recent approaches leverage appearance features and incorporate temporal information to improve discriminative power for better occlusion handling. However, most existing methods fail to fully utilize historical tracking information without significantly compromising tracking efficiency. To address this limitation, we propose a new MOT framework called the feature-adaptive continual learning tracker (FACT), which supports online learning while effectively leveraging the complete tracking history. At the core of this framework is the feature-adaptive continual-learning (FAC) module, which adopts an analytic continual learning to enable online training using complete historical tracking information. Specifically, FAC module updates appearance models using only current-frame features and a fixed-size feature auto-correlation matrix via closed-form update rules, achieving performance equal to joint training on all past tracking data. We show that the FAC module can be seamlessly integrated into existing feature-based trackers, enhancing performance with minimal impact on tracking speed. Additionally, we introduce a two-stage association mechanism tailored to the proposed analytic continual learning-based paradigm. Extensive experiment results demonstrate that the proposed method achieves state-of-the-art online tracking performance on MOT16, MOT17, and MOT20 benchmarks.

Accurate and robust energy consumption forecasting across heterogeneous spatial zones remains a critical challenge for modern smart grids [7][15] Power demand exhibits hierarchical temporal structure-diurnal trends, intrahour stochasticity, and zone-specific anomalies-yet existing single-pass forecasters conflate these hierarchies into monolithic architectures, inevitably compromising performance on at least one scale. We propose HTFSN (Hierarchical Temporal Feature Sedimentation Network), a cascade framework that explicitly decomposes forecasting into two complementary hierarchies: (1) Trend Sedimentation via PatchTST extracts coarse-grained global patterns through patch tokenization and transformer attention, and (2) Residual Sedimentation via N-HiTS refines zone-specific, fine-grained patterns through multi-scale dilated convolutions. We further introduce zone-adaptive residual fusion-a cross-validated per-zone $\alpha$-optimization-and validation-gated ensemble integration with strict improvement thresholds preventing overfitting. On real-world smart meter data (20,000 samples, 12-step forecasting), HTFSN achieves 29.4% MSE reduction $(\mathbf{5 . 1 3 M} \rightarrow \mathbf{3 . 6 2 M}), \mathbf{2 0 . 1 \%}$ MAE improvement $(\mathbf{1, 7 1 3} \rightarrow 1,369)$, and $\mathbf{R}^{2}=0.850$ compared to single-stage PatchTST. Comprehensive ablations, horizon-wise decomposition, error distributions, and parity analysis validate each component. PatchTST’s $4.2-7.2 \times$ superiority over classical architectures (Transformer / LSTM / TCN) establishes it as an optimal trend extractor, further refined by our residual sedimentation approach.

Latent domain watermarking for diffusion models embeds watermarks directly into the latent prior, enjoying non-intrusiveness to model parameters and seamless integration with the generation process. However, due to the violation of latent Gaussianity or sensitivity to normal and malicious perturbations during latent inversion, existing methods are prone to watermark detection or removal attacks. A further overlooked problem is the violation of the i.i.d. latent condition after watermarking, which leads to latent correlation degradation and generation fidelity loss. Although this has been externally measured by FID, the internal correlation structure has yet to be rigorously characterized. To address the above issues, and motivated by the rotation-invariant property of isotropic Gaussian, we propose Latent Angular Watermarking (LAW), which encodes watermark bits as antipodal angles (/2 relative to a reference pair) between disjoint pairs of latent elements while preserving the Gaussianity. The antipodal (π-separation) encoding maximizes geometric separation between bit values, and we prove that the decoding angular-error variance is proportional to the norm of the latent pair, i.e., var(Δϕ) ∝ 1/ρ^2. We further propose a magnitude-driven variant, LAW-M, which anchors watermark bits in the most geometrically stable latent dimensions, yielding additional robustness gains. Theoretically, we provide a rigorous characterization of the induced correlation degradation, deriving in closed form the autocorrelation structure of the watermarked latent and proving that correlations are confined to a sparse, structured set of off-diagonal elements with fixed /4 values.

A direct digital frequency synthesizer (DDFS) which requires high spurious-free dynamic range (SFDR) performance often faces significant hardware overhead, particularly in its reliance on dedicated multipliers within digital signal processing (DSP) slices or large lookup tables (LUTs) within block RAMs (BRAMs). This paper presents a multiplier-less, high-SFDR DDFS architecture that incorporates a novel 3-segment Adaptive Recoding CORDIC (ARC) design to improve convergence and SFDR under multiplier-less constraints. Furthermore, an Adaptive Upward Rounding Algorithm (AURA) is introduced as an empirically effective sub-LSB error compensation strategy, motivated by statistical error characteristics, which helps suppress truncation-related spurs with negligible hardware overhead. Implementation on an AMD Artix-7 FPGA demonstrates that the 17-bit high-precision model achieves an average SFDR of 119.65 dBc at a maximum clock frequency of 259.7 MHz, while the 15-bit resource-efficient variant provides 106.84 dBc SFDR at 269.5 MHz, both without using any DSP slices or BRAMs.

Time series classification is crucial in many real-world applications, but models often suffer significant performance degradation under domain shifts between training and testing data. Existing Domain Generalization (DG) methods for time series mainly suppress domain-specific information via adversarial training, but they remain insufficient to obtain category-consistent representations. To address these challenges, we propose Joint Inference for Time Series Representation (JITSR), a variationally motivated novel framework for domain-generalized time series classification. Through joint variational modeling of time series and labels, JITSR connects patch reconstruction, categorical prediction, and distributional alignment as complementary learning objectives. The alignment term encourages category-consistent representations across domains without requiring explicit domain annotations. Specifically, JITSR segments time series into patches and maps them to latent distributions, which are quantized by a learnable distributional codebook. A dual-path alignment mechanism then aligns sample-level semantic distributions with codebook-level class priors. Meanwhile, attention-based codeword weighting identifies the temporal patterns that contribute to classification, providing interpretable evidence for the learned representations. Extensive experiments on diverse time series datasets demonstrate that JITSR consistently outperforms state-of-the-art DG methods and learns more category-consistent representations across domains.

Accurate object tracking holds significant importance in applications where a specific object needs to be tracked amidst many visually similar objects in complex environments, such as scenarios involving occlusion and low-light conditions. Considering the efficiency, convenience, and safety of ultrahigh-frequency (UHF) radio frequency identification (RFID), this work explores the use of a robot for dynamic tag tracking, where tags are attached to mobile objects. It is the first study to develop a purely RFID-based localization technique which enables the mobile robot to localize moving tags using on-board antennas by introducing virtual candidate positions, which seamlessly integrates the RFID range capabilities with the directional focus of the antenna array. Due to the limited sensing capabilities of RFID, a novel robot tracking method is proposed, which incorporates RFID detection constraints into the design process of an adaptive controller. The stability of the proposed method is then analyzed rigorously. Experiment results in different scenarios (normal scene, dark scene, and occlusion scene) demonstrate the average positioning error is about 0.03 meters, which is better than state-of-the-art technologies. In addition, the proposed algorithm can track specific targets concurrently in environments with similarities, occlusions, and multiple tags.

The language modeling paradigm for scene text recognition (STR) has demonstrated impressive universal capabilities across extensive STR scenarios. However, existing methods still encounter challenges in effectively handling text images with irregular shapes and diverse appearances (e.g., curve, artistic, multi-oriented) due to the absence of contextual information during initial decoding. In this work, inspired by the principle of ‘forest before trees’ in human visual perception, we introduce NASTR, a non-autoregressive scene text recognizer capable of endowing global-aware for the attentional decoder. Specifically, we design a global-to-local attention procedure, simulating the mechanism of globally holistic visual signal processing preceding locally detailed response in the human brain visual system. This is achieved by leveraging the global image information queries to condition the generation of glimpse vectors at each decoding time step. This procedure empowers the NASTR model to achieve on-par performance with its state-of-the-art autoregressive counterparts, while operating in a fully parallel manner. Moreover, we propose multiple optional and flexible encoding constraint components to alleviate the representation quality degradation issue caused by the global image information queries in handling STR tasks with multilingual and in multi-domains. These components constrain the global image features from the perspective of global structural, global semantic, and linguistic knowledge. Extensive experimental results demonstrate that NASTR consistently outperforms existing methods on both Chinese and English STR benchmarks. Our source code, trained models, and logs are available at https://github.com/ML-HDU/NASTR.

Energy consumption forecasting requires balancing model sophistication with practical deployability, particularly under extreme weather volatility and data scarcity. This paper proposes a two-stage hybrid framework integrating advanced temporal modeling with generative data augmentation.In Stage 1, a Ensemble architecture combines Transformer (for long-range dependencies), LSTM (for sequential dynamics), and TCN (for multi-scale patterns) networks, fused by an XGBoost meta-learner. A key innovation is the dual-pipeline design that simultaneously forecasts daily maximum and minimum temperatures, precisely capturing diurnal thermal amplitude essential for peak load planning.To address extreme-weather sample scarcity, a three-phase GAN framework is introduced, integrating RGANomaly for anomaly detection, a GASN module with Multi-Head Self-Attention for latent temporal modeling, and TAnoGAN to synthesize realistic extreme-temperature scenarios.In Stage 2, these enhanced forecasts drive multivariate consumption models (SARIMAX, VARMAX) across heterogeneous sectors. Results confirm that modeling inter-sector dependencies is critical, reducing Service sector prediction error by 18.6%. The proposed framework achieves a 9.8% reduction in combined forecasting error compared to a robust XGBoost baseline(1024 estimators). Validated on five years of real-world data, the method demonstrates superior robustness in capturing high-volatility events while remaining deployable in resource-constrained utility environments.

Multi-modal image fusion aims to integrate complementary cues from different modalities into a single image, facilitating downstream tasks such as object detection. However, input image pairs are often misaligned due to rigid or non-rigid deformations in practical scenarios. Such misregistration introduces structural distortions and visual artifacts, reducing the reliability of the fused results and limiting their effectiveness for downstream vision applications. While existing methods demonstrate satisfactory results under specific deformation scenarios, they exhibit limited generalization to diverse and severe misregistrations. To this end, this study proposes a unified multiple-in-one flow Mamba framework for registering various image deformations and generating high-quality fused results. Specifically, a hierarchical flow Mamba is designed to model rigid and non-rigid flow fields and enhance adaptability to complex deformations by progressively refining misaligned features. To better distinguish between rigid and non-rigid types, a flow field classifier predicts rigid/non-rigid categories and provides prompts for high-level feature modulation. Furthermore, a flow-guided fusion Mamba module is developed to aggregate aligned multi-level modality features and generate fused images, while an iterative training strategy enables collaborative optimization by using fusion outputs to refine flow estimation. Experiments across three representative modality tasks demonstrate that the proposed method delivers superior fusion performance while maintaining applicability to object detection. The code will be available at: https://github.com/BOYang-pro/MOFM.

Unsupervised cross-modal hashing enables efficient retrieval of semantically related instances across different modalities without requiring manual semantic annotation. However, existing unsupervised methods rely heavily on large-scale image-text pairs. Collecting such data can be costly, particularly in scenarios where well-aligned pairs are scarce due to privacy and specialized constraints. More critically, existing methods tend to overfit to seen training data, restricting their generalization performance on unseen categories that the constrained training data cannot cover. To address these limitations, we propose Attribute-Prompted Kernel Hashing (APKH), a novel data-efficient approach that constructs a compact, modality-aligned Hamming space driven by the generalized attribute priors of vision-language foundation models. Specifically, APKH introduces two core modules: Context-optimized Attribute Kernel Mapping (CAKM) and Kernel-Smoothed Contrastive Alignment (KSCA). CAKM formulates cross-modal alignment through hyperspherical Radial Basis Function kernel mapping, optimizing dynamic attribute kernels via prompt learning to capture modality-invariant semantics. Furthermore, KSCA extends conventional point-to-point contrastive learning by modeling limited paired data as continuous kernel distributions. This explicit smoothing of the modality gap alleviates overfitting to sparse pairwise correlations. Extensive experiments demonstrate that APKH outperforms state-of-the-art hashing methods in the challenging cross-modal retrieval tasks from seen to unseen categories under data-constrained scenarios.

Current research on integrated sensing and commu nication (ISAC) primarily examines the use of communication waveforms to derive radar parameters, but there is a significant gap in methods for accurately locating targets using these parameters. Therefore, this paper explores target localization solutions within an ISAC system under two scenarios. First, we investigate tracking a moving target using multistatic sensing with dynamic transmitters (TXs). Unlike existing two-step weighted least square (2SWLS) methods, which rely on the bistatic range (BR) and bistatic range rate (BRR) measurements, our method incorporates an additional direction of arrival (DOA) measurement. This innovation reduces the minimum number of required transmitter receiver (TX-RX) pairs, needing only one TX-RX pair for position estimation and two TX-RX pairs for velocity estimation, thereby improving system efficiency and accuracy. In the second scenario, we address the challenge of unknown-location TXs. We pioneer a novel method that simultaneously tracks the moving target and determines the locations of these unknown-location TXs. This approach operates effectively with only three TX-RX pairs, overcoming the significant obstacle of operating without prior knowledge of TX positions. Simulations validate the effectiveness of our proposed methods. For the first scenario, our method demonstrates superior performance compared to existing 2SWLS methods. For the second scenario, our method successfully navigates the complexity of unknown-location TXs, achieving reliable target tracking and TX localization. As a result, this paper fills the gap in current ISAC studies by providing detailed methods for accurately locating a target using radar parameters obtained from an ISAC receiver.

Box-free model watermarking is an emerging technique to safeguard the intellectual property of deep learning models, particularly those for low-level image processing tasks. Existing works have verified and improved its effectiveness in several aspects. However, in this paper, we systematically investigate the vulnerability and demonstrate that box-free model watermarking is prone to removal attacks, even under the real-world threat model such that the protected model and the watermark extractor are in black boxes. Under this setting, we carry out three studies. 1) We develop an extractor-gradient-guided (EGG) remover and show its effectiveness when the extractor uses ReLU activation only. 2) More generally, for an unknown extractor, we leverage adversarial attacks and design the EGG remover based on the estimated gradients. 3) Under the most stringent condition that the extractor is inaccessible, we design a transferable remover based on a set of private proxy models. In all cases, the proposed removers can successfully remove embedded watermarks while preserving the quality of the processed images, and we also demonstrate that the EGG remover can even replace the watermarks. Extensive experimental results verify the effectiveness and generalizability of the proposed attacks, revealing the vulnerabilities of the existing box-free methods and calling for further research.

This paper introduces the Reservoir-Augmented Asymmetric (RAA) Kernel, a novel approach for efficient and accurate Human Activity Recognition (HAR) using WiFi Channel State Information (CSI). While traditional ma chine learning methods often fail to capture the temporal dynamics inherent in CSI data, deep learning approaches demonstrate better performance but suffer from high computational demands. To address these issues, we lever age the structural advantages of Reservoir Computing (RC) to effectively capture the time series dynamics, and generate high-quality interpretable features from the CSI data. We further enhance this model by incorporating Gaussian noise within a unique asymmetric kernel framework, mitigating overfitting without increasing computa tional complexity. The RAA kernel's effectiveness is demonstrated through extensive experimentation on various CSI-based HAR datasets, showing notable improvements over both traditional and deep learning approaches. The source code is publicly available at https://github.com/linghan1997/Reservoir-augmented-asymmetric-kernel-for-WiFi-based-human-activity-recognition.

Time series modeling faces two challenges: out-of-distribution (OOD) generalization to correctly classify sequences from distribution shift domains, and OOD detection to identify anomalous or previously unseen sequences. Existing methods usually address these separately, which limits their applicability in real-world scenarios. In this work, we propose UniOOD, a unified framework that tackles both challenges simultaneously. UniOOD divides sequences into patches, encodes each as a latent Gaussian distribution, and aggregates them into a sequence-level posterior via learned weighting. By aligning sequence posteriors with learnable class priors while maintaining inter-class separation, UniOOD learns representations that are both generalizable and discriminative. The distance between posteriors and class priors provides an intrinsic OOD score, enabling detection of unseen sequences without prior exposure. Experiments show UniOOD achieves state-of-the-art classification, strong OOD generalization, and effective OOD detection within a single unified model.

The cycloid gear bearing (CGB) in rotate vector (RV) reducers undertakes heavy torque load cyclically, and is more prone to fault. However, CGBs feature involute kinematics, complex dynamics and weak fault features in vibrations. This poses a tremendous challenge to traditional vibration based fault diagnosis methods, which are confined by sensor placement issue and vibration complexity. To address this issue, the merits of motor stator current measurement and analysis are leveraged, such as measurement convenience and lower signal complexity. For this purpose, the calculation equation of CGB fault characteristic frequency is derived, based on kinematics analysis of RV reducer and CGB. Furthermore, the motor current modulation mechanism and spectral characteristics under CGB fault are elucidated analytically. Specifically, the amplitude modulation and frequency modulation model of motor stator currents are derived firstly, through kinetics analysis of RV reducer CGBs and electrical-magneto-mechanical interaction analysis of electric drive motor - RV reducer drivetrain. Such analytical models reveal the multiple modulation nature of motor current oscillation, and motivate demodulation analysis idea for CGB fault characteristic frequency identification. Then, the explicit equations of amplitude and frequency demodulation spectra are deduced, based on the derived analytical models. These derivations justify the presence of fault characteristic frequency, cycloid and cage rotating frequencies from the kinetics perspective, and discover the CGB fault spectral symptom law manifested in motor currents. Finally, a joint demodulation methodology is proposed for reliable and complementary fault feature extraction, based on amplitude and frequency demodulation analysis. The theoretical derivations are validated experimentally through lab test data analysis of an RV reducer. Localized fault on the inner race, outer race, and needle roller of CGBs are all successfully diagnosed.

Passive Human Activity Recognition (HAR) utilizing ambient wireless signals is becoming a cornerstone of privacy-preserving healthcare and smart environments. However, current solutions predominantly rely on controlled WiFi setups or dedicated transmitters, which restricts their scalability and real-world coverage. In contrast, commercial 5G base stations (BSs) offer pervasive signal availability but introduce significant challenges, notably low Signal-to-Noise Ratio (SNR) and synchronization errors inherent in non-cooperative reception. To capitalize on the extensive reach of modern cellular networks, this paper presents a novel HAR system that harvests Channel State Information (CSI) directly from commercial 5G downlink broadcast signals which may contain some unintended side channel information (SCI). To address the inherent instability of raw commercial signals, we propose a robust channel-centric representation framework. This approach integrates latent representations with time-domain channel features and captures their temporal dynamics through a learning-based architecture. Extensive experiments in a live commercial 5 G environment under challenging low-SNR conditions demonstrate a classification accuracy of 90.11 % across four activity classes, effectively repurposing existing 5 G infrastructure into a ubiquitous sensing network for smart environments and cyber-security applications.

Compared to supervised cross-modal hashing (CMH), unsupervised CMH reduces the reliance on manual labeling by learning binary codes from unlabeled image-text pairs. However, existing unsupervised CMH methods often rely on large-scale image-text pairs, which are costly to collect. To address this limitation, we propose Global-Neighborhood Alignment Hashing (GNAH), a novel approach that preserves the semantic structure of vision-language foundation models within a compact binary Hamming space using only a limited number of image-text pairs. Specifically, GNAH captures global structural information from the continuous latent space and transfers it into the binary Hamming space through a Prototype-Anchored Global Alignment module. In addition, GNAH extends conventional pairwise contrastive learning by modeling stochastic neighborhood relationships via a Contrastive Stochastic Neighborhood Alignment module, thereby alleviating overfitting to sparse pairwise correlations. Extensive experiments demonstrate that GNAH consistently outperforms existing unsupervised cross-modal retrieval methods under data-constrained settings, offering a practical solution for real-world CMH applications.

In light of their capability to capture structural information while reducing computing complexity, anchor graph-based multi-view clustering (AGMC) methods have attracted considerable attention in large-scale clustering problems. Nevertheless, existing AGMC methods still face the following two issues: 1) They directly embedded diverse anchor graphs into a consensus anchor graph (CAG), and hence ignore redundant information and numerous noises contained in these anchor graphs, leading to a decrease in clustering effectiveness; 2) They drop effectiveness and efficiency due to independent post-processing to acquire clustering indicators. To overcome the aforementioned issues, we deliver a novel one-step multi-view clustering method with adaptive low-rank anchor-graph learning (OMCAL). To construct a high-quality CAG, OMCAL provides a nuclear norm-based adaptive CAG learning model against information redundancy and noise interference. Then, to boost clustering effectiveness and efficiency substantially, we incorporate category indicator acquisition and CAG learning into a unified framework. Numerous studies conducted on ordinary and large-scale datasets indicate that OMCAL outperforms existing state-of-the-art methods in terms of clustering effectiveness and efficiency.