Search Results (238)

Pedestrian safety at urban intersections remains a major challenge in Intelligent Transportation Systems (ITSs). This study investigates whether crossing intention can be reliably inferred directly from temporal body-pose dynamics to drive real-time collision warnings on embedded edge platforms. Existing vision-based approaches that rely primarily on bounding-box proximity or scene-level spatial grids are often prone to false alarms in complex urban environments with motorcycles, stationary pedestrians, and background clutter. To overcome these limitations, we propose an end-to-end framework consisting of four sequential processing stages: (1) a perception layer integrating YOLOv8s, ByteTrack, a displacement filter, and rider suppression to generate reliable pedestrian trajectories; (2) a skeleton extraction layer utilizing YOLOv8s-pose to construct temporal sequences of 17 anatomical keypoints; (3) an ultra-lightweight Skeleton Graph Convolutional Network (SkeletonGCN, comprising 33.8 K parameters, <0.2 MB) that models body-joint kinematics and temporal motion dynamics; and (4) an image-space Time-to-Collision (TTC) risk-fusion module. While this fusion approach avoids explicit geometric camera calibration, it still relies on predefined scene-profile parameters and image-space motion assumptions. Furthermore, while the intention classifier is quantitatively evaluated, the risk-fusion module is procedurally defined, and its resulting four-level collision warnings are demonstrated operationally rather than validated against ground-truth hazard annotations. Evaluated on 49,948 valid sequences from the JAAD and PIE benchmark datasets under a strict video-level partitioning protocol, the unified SkeletonGCN achieves a macro-F1 score of 0.717 (with per-scene subset macro-F1 scores of 0.761 on JAAD/PIE urban and 0.895 on intersections), significantly outperforming baseline models. When deployed on an NVIDIA Jetson Orin NX edge device using TensorRT FP16, the full pipeline achieves an instrumented latency of 70.7 ms per frame (~14 fps) and a sustained wall-clock throughput of 7.4 fps on real-world urban dashcam video. System limitations include sensitivity to 2D printed human imagery and reduced prediction reliability under low-light nighttime conditions. Full article

To address the challenges in multi-modal trajectory prediction for multi-aircraft interactions within non-towered terminal airspace, including the insufficient extraction of long-range temporal dependencies, neglect of physical separation constraints, and barriers to integrating flight intentions and multi-source environmental context, this paper develops a trajectory prediction model integrated with long-range temporal modeling, physics-aware spatial interaction, and intention context enhancement. A parameter-shared ST-Transformer temporal encoder is established to capture long-term motion patterns of aircraft via global multi-head self-attention, and temporal attention pooling is adopted to mitigate error accumulation in long-term prediction. A physical distance-aware ST-GAT module is designed, which embeds the spatial distance prior between aircraft into the attention calculation and leverages distance masks to reduce interference from distant irrelevant aircraft. Furthermore, an intention-aware context enhancement module (IACEM) is proposed. It identifies the distribution of flight phases and adaptively incorporates meteorological information to construct enhanced features embedded with high-level semantics and environmental priors. Finally, a CVAE-based framework is utilized to generate multiple candidate trajectories satisfying kinematic constraints. Multiple verification experiments are carried out on the TrajAir dataset. The experimental results demonstrate that the proposed model outperforms various baseline models in terms of ADE and FDE. Ablation studies and visual analysis verify that the three core modules produce synergistic improvements. The model achieves higher prediction accuracy under scenarios involving 2D complex maneuvers, 3D climbing turns, and dense multi-aircraft interactions, which proves the effectiveness and superiority of the proposed algorithm for trajectory prediction in non-towered terminal airspace. Full article

As an emerging sensing modality, 4D millimeter-wave radar provides a privacy-preserving and lighting-robust solution for indoor scene perception and human action recognition (HAR). However, existing radar-based HAR methods mainly focus on short-term motion dynamics while overlooking the influence of indoor scene context. Furthermore, reconstructing indoor layouts from sparse and irregular radar point clouds remains challenging. To address these issues, this paper proposes an environmental context-aware radar action recognition framework (ECRAR) that jointly models human motion and indoor scene context from 4D millimeter-wave radar point clouds. Specifically, a trajectory-to-floorplan inversion branch reconstructs indoor layouts from long-term trajectory point clouds by exploiting the relationship between human movement distributions and spatial accessibility. A motion feature extraction branch captures discriminative short-term motion representations through multi-scale spatial modeling, while an environmental context fusion module adaptively integrates scene and motion features using cross-attention. We also construct InActivity-Scene, a new dataset with fine-grained scene annotations and nine categories of daily and hazardous human actions collected in multiple indoor environments. Experimental results demonstrate that ECRAR consistently outperforms existing methods, achieving 92.2% recognition accuracy with significant improvements on environment-dependent actions, while ablation studies further verify the effectiveness of each proposed component. Full article

Subway tunnels exacerbate fire hazards, revealing a critical lacuna regarding the spatio-temporal coupling of fire progression and human egress. To address this, this study proposes and validates an Integrated Spatio-Temporal Risk Assessment Model to explicitly quantify survivability thresholds under complex fire dynamics. The framework synergizes Large-Eddy Simulation (LES) based Computational Fluid Dynamics with agent-based pedestrian trajectory modeling within a 3D tunnel featuring a 2% longitudinal gradient. Evaluating 9.5 MW and 12 MW fire energies, the model assessed Single-Sided Evacuation (SSE), Double-Sided Evacuation (DSE), and Sprinkler-Assisted Single-Sided Evacuation (SSE-S) across 2160 agents. Hazards were quantified by continuously resolving Fractional Effective Dose (FED) indices, 60 °C boundaries, and 500 ppm CO fronts. Simulations reveal the gradient induces a severe stack effect, accelerating toxic dispersion and yielding temperatures exceeding 1200 °C. Consequently, SSE engendered fatal bottlenecks (FED: 15.33), whereas DSE optimized pedestrian flux, capping peak FED at 0.52. Crucially, while active suppression (SSE-S) extinguished flames within 105 s, thermodynamic cooling induced a paradoxical loss of smoke buoyancy, causing toxic layers to stratify at the breathing zone. Ultimately, while DSE and SSE-S are paramount for survivability, water-based suppression generates localized toxicological risks, necessitating the integration of low-level smoke detection and extraction architectures in future subterranean designs. Full article

Urban rail transit infrastructure inspection increasingly relies on Light Detection and Ranging (LiDAR) due to its capability for efficient and high-precision 3D data acquisition. However, robust rail tread segmentation in ballastless metro environments remains challenging due to boundary leakage, interference from geometrically similar structures, and the heavy dependence of existing methods on Red-Green-Blue (RGB) imagery, trajectory priors, or template matching. To address these limitations, this study proposes a lightweight rail tread extraction framework for ballastless track LiDAR point clouds based on multi-stage filtering and curvature-guided region growing. First, intensity thresholding and cloth simulation filtering are leveraged to prune tunnel walls, track beds, and other large-scale non-target structures, thereby reducing computational overhead. Subsequently, local normal vectors and curvature features are estimated via Principal Component Analysis (PCA). A curvature-ranked seed selection strategy and a dual-constrained region growing mechanism, integrating normal consistency and curvature thresholds, are then introduced to suppress excessive growth near rail boundaries and enhance regional homogeneity. Experimental results on field data collected from Shanghai Metro Line 10 demonstrate that the proposed method achieves a recall of 92.23%, a precision of 95.32%, and an F1-score of 93.7%, outperforming conventional Euclidean clustering and standard region growing algorithms. Compared with deep learning approaches, the proposed framework requires no large-scale annotated training data and is independent of RGB information or trajectory priors, making it better suited for lightweight engineering deployment in practical urban rail transit maintenance. Full article

To address sustained deviation from coverage lanes when an unmanned surface vehicle (USV) executes a complete boustrophedon coverage path under current disturbances, this study proposes an adaptive current-aware line-of-sight (ACA-LOS) guidance method based on online current estimation. The local current is estimated online from the kinematic velocity relationship between the USV and the surrounding water, and its path-normal component is extracted. The estimated path-normal current is used both to construct a current-compensating heading correction and, together with the cross-track error, to adjust the LOS lookahead distance and generate the desired ACA-LOS heading. Complete-path simulations are conducted under constant, time-varying, and spatially varying currents with different current speeds and directions. Relative to conventional LOS, ACA-LOS reduces the mean cross-track root mean square error (RMSE) by 40.80%, increases the tolerance-band compliance rate ($R_b$) by 27.48 percentage points, and reduces the cumulative cross-track deviation area ($A_{\mathrm{d}}$) by 68.68%. ACA-LOS suppresses sustained lane deviation caused by current disturbances and improves lane keeping throughout complete boustrophedon path execution. The method supports the extension of USV operation from single-trajectory tracking to long-duration, continuous coverage execution and provides a reference for unmanned and intelligent waterborne operational equipment. Full article

Long-horizon robotic manipulation requires a policy to bridge task-level semantic reasoning with metric three-dimensional interaction geometry. Existing vision–language–action policies usually acquire geometry implicitly from visual tokens or introduce deterministic intermediate variables only in the image plane, which rely on expensive human annotations and training cost. This article presents a spatial Gaussian-guided hierarchical framework that uses ordered 3D Gaussian interaction regions as an explicit planning interface between vision–language reasoning and action generation. The proposed framework enables efficient adaptation of a pretrained vision–language model for robotic manipulation tasks. First, an automatic geometric enhancement pipeline converts raw robot demonstration videos into near-, mid-, and late-stage Gaussian supervision through foreground extraction, metric depth estimation, stable camera aggregation, end-effector localization, 3D lifting, and temporal grouping, without requiring manual 3D interaction annotation. The generated Gaussian representations provide structured spatial guidance, where their covariance characterizes interaction-region extent and variability rather than fully calibrated physical uncertainty. Second, a shared vision–language backbone predicts structured subtasks and Gaussian interaction regions, while a conditional diffusion executor generates future action chunks under these semantic and geometric conditions. A trajectory-to-Gaussian likelihood objective explicitly encourages consistency between generated motions and the predicted spatial interaction plan. Experiments on a mixed real-robot dataset derived from LHManip and RH20T show that our method improves trajectory tracking success from 55.7% to 70.8% over a same-backbone direct VLA baseline. Closed-loop simulation evaluation on LIBERO with 80% backbone parameter frozen achieves 85.3% average task success, demonstrating the effectiveness of explicit 3D interaction representations for spatial reasoning and long-horizon manipulation. Full article

Accurate State-of-Health (SOH) prediction is essential for deploying retired electric vehicle batteries into reliable second-life energy storage systems. However, this task is challenged by sparse and noisy operational data from onboard Battery Management Systems (BMS). This study systematically evaluates a spectrum of deep learning architectures for SOH forecasting under BMS-style data constraints derived from laboratory cycling data: a BiLSTM on aggregated cycle statistics (Model A), preliminary zero-shot transfer to a single unseen cell (Model B), a waveform BiLSTM with full intra-cycle voltage, current, and temperature trajectories (Model C), a baseline TCN (Model D) and a probabilistic TCN-GPR hybrid (Model E). All models are constrained to identical low-fidelity BMS-style variables extracted from the NASA battery aging dataset. Model C achieves the lowest point accuracy error of 0.46% ± 0.18% MAE across five random seeds, demonstrating that high-resolution waveform inputs capture degradation signatures, notably voltage plateau morphology, transient dynamics, and implicit SOC information, that aggregated features irreversibly lose. Model D using the same waveform inputs and evaluation protocol as Model C, achieves a MAE of 2.99% at a single seed, providing direct architectural comparison evidence that the BiLSTM’s position-sensitive temporal summarization outperforms GlobalAveragePooling1D under these conditions. Model E achieves a higher MAE of 2.12% ± 0.33% but uniquely provides calibrated predictive distributions of 99.4% ± 1.2% coverage, NLL = −1.877 ± 0.038, with approximately uniform 95% predictive intervals (mean width 19.83% SOH across 34 test cycles at seed = 42), reflecting the near-constant posterior variance produced by the large optimized GPR length-scale under the frozen two-stage training design. A paired t-test confirms that Model C statistically significantly outperforms Model E on point accuracy (p < 0.01). Isotonic regression recalibration reduces mean calibration error from 0.138 to 0.010, demonstrating that shape-level miscalibration is correctable post hoc. The central implication for second-life battery deployment is a clear accuracy–uncertainty trade-off: Model C is preferred when point estimates suffice, while Model E is essential for risk-aware decisions requiring confidence intervals. Full article

This study presents a sensor-fused field investigation and simulation-based analysis of four horizontal and vertical traffic-calming devices—two raised speed tables, a speed hump, and a raised crosswalk—installed along a 5250-ft two-lane residential collector in Nashville, TN, USA. A dual-sensor architecture combining a Miovision Scout video-based vehicle counter and WAAS/EGNOS-augmented GPS probe-vehicle logging (5 m 3-D RMS horizontal accuracy, 1 Hz sampling) was used to reconstruct 30 quality-controlled free-flow vehicle trajectories and 12-h per-lane volume counts. A spatial kinematic transform (a = v·dv/dx) was applied to extract device-specific approach-deceleration and post-device recovery-acceleration rates, and a three-parameter log-logistic cumulative-distribution function was fitted to the field-observed desired-speed percentiles (root-mean-square error below 0.043 for both speed-table devices). The camera- and GPS-derived observations were used to calibrate and statistically validate a PTV VISSIM microsimulation replica of the corridor, achieving a mean-speed calibration error of 0.71% or better at every device, a GEH statistic below 1.5 at all four analysis turning movements, and independent travel-time validation errors of 5.7–12.1%, within the accepted 15% threshold. The validated model was then used to reconstruct device- and spacing-specific May–Keller macroscopic speed–density–flow relationships, calibrated against simulated capacities of 650–775 vehicles per hour per lane at 350-, 700-, and 1050-ft device spacing. Results show capacity reductions of 20–33% relative to free-flow conditions and yield kinematically derived maximum recommended spacings of 265–630 ft to maintain crossing speeds at or below 15 mph, depending on device geometry. The findings demonstrate a reproducible, low-cost sensor-fusion workflow for quantifying the safety–capacity trade-off of traffic-calming corridors and for informing the design of sensor-in-the-loop adaptive-calming infrastructure. Full article

Cervical range of motion (CROM) has been used in research and clinically for assessing cervical health. Gold-standard goniometers tend to be cumbersome. However, Inertial Measurement Units (IMUs) or vision-based alternatives demand frequent calibration and/or costly hardware; moreover, the subject is aware of being measured and there is a risk of breaching privacy. In response, we have developed a non-contact HPNet system for head pose estimation (HPE) that can use a rear-facing camera to quantify CROM accurately. A Re-parameterized Visual Geometry Group (RepVGG)-D2se model is employed as the backbone of the network, and a Spatial Feature Enhancement (SCFE) module is incorporated to improve feature extraction. HPNet was evaluated on the large-scale Carnegie Mellon University (CMU) Panoptic dataset, achieving a mean absolute error (MAE) of 3.48°, 3.22°and 3.34° for yaw, pitch and roll respectively. Inter-instrument reliability was excellent for all six cervical movements when compared with the research/clinical-grade CROM device, with intraclass correlation coefficients (ICCs) averaging 0.939. Bland–Altman plots confirmed close agreement between the two methods. Cervical movement trajectory curves further confirmed the concordance between the clinical device and our method. The system is fully automatic, requires only a rear-facing camera, effectively preserves patient privacy, and provides accurate cervical posture estimation. This technology may provide a basis for future applications in neck-disorder screening, remote health monitoring, and personalized musculoskeletal wellness management, although further task-specific clinical validation will be required. To date, HPNet has been validated primarily on a computer-based platform and has not yet been deployed on smartphones. Future work will focus on model lightweighting, mobile deployment, and cross-device adaptation to facilitate its practical implementation on mobile devices. Full article

Background/Objectives: Allergic rhinoconjunctivitis (AR) induced by house dust mites (HDMs) is a common condition that impairs quality of life, and allergen immunotherapy is the only treatment accepted to modify its natural course. However, real-world evidence on polymerized HDM subcutaneous immunotherapy (SCIT) remains limited. We evaluated the up to 24-month effectiveness of polymerized HDM SCIT and examined how baseline symptom burden affects the measured response. Methods: In a retrospective, single-arm, real-world cohort, 353 patients (mean age 15.9 (12.3) years) with HDM-induced AR with/without asthma (asthma subpopulation, n = 112) initiated polymerized SCIT (ALXOID^®, Inmunotek S.L., Spain) with an extract of house dust mites (Dermatophagoides pteronyssinus, D. farinae, Blomia tropicalis). Patients were followed for up to 24 months; 139 contributed an observation at month 12 and 67 at month 24. The primary outcome was the Rhinoconjunctivitis Combined Symptom–Medication Score (RCSMS); secondary outcomes were a visual analog scale (VAS) and the ESPRINT-15 quality-of-life questionnaire. Effectiveness was estimated using mixed models for repeated measures. The pre-specified primary-analysis population was restricted to a baseline RCSMS ≥ 2 to avoid floor effects and regression to the mean at low baseline scores. Results: At month 24, the RCSMS decreased by 0.938 points (−35.3%) in the full cohort and by 48.3% in the baseline RCSMS ≥ 2 population; in the full cohort, the VAS and ESPRINT-15 improved by 50.6% and 69.7%, respectively (all p < 0.001; within-subject effect sizes were small for the full-cohort RCSMS and medium for all other outcomes). The shape of the improvement trajectory did not differ by age group, and all age groups exceeded the minimal clinically important difference; the age analysis was exploratory. Patients with baseline RCSMS < 2 showed continued improvement in VAS and ESPRINT-15 scores. In contrast, the RCSMS increased, a pattern consistent with a statistical floor effect, suggesting limited responsiveness of the score at low baseline values. Conclusions: Up to 24 months, polymerized HDM SCIT was associated with clinically meaningful improvements in symptoms, medication use, and quality of life, most clearly in patients with a baseline RCSMS ≥ 2. Baseline symptom severity should be considered when interpreting combined symptom–medication outcomes in allergen immunotherapy. The single-arm design precludes causal inference. Full article

Temperature–time profiles obtained through thermistor-based monitoring provide a rich but noise-sensitive source of information for estimating kinetic and thermal parameters in exothermic batch reactions. Conventional workflows typically combine deterministic smoothing with numerical differentiation, an approach that can amplify measurement noise and fail to propagate preprocessing uncertainty into the resulting reaction-rate and parameter estimates. To address these limitations, this study presents an uncertainty-aware Bayesian machine-learning framework that integrates scalable random-Fourier-feature Gaussian-process (RFF–GP) smoothing, analytical differentiation, temperature-derived apparent conversion, Bayesian parameter inference, posterior validation, predictive calibration, model comparison, ablation, sensitivity analysis, probabilistic benchmarking, simulation of thermal nonideality, and endpoint diagnostics. The framework was applied to $379,631$ cleaned thermistor observations. The production RFF–GP achieved a validation root-mean-square error of $0.04805\mathrm{K}$, yielding a stable latent temperature trajectory and an uncertainty-aware estimate of $dT/dt$. On a smaller matched subset, exact Gaussian-process regression achieved the highest predictive accuracy and the best probabilistic scores, whereas the RFF–GP reduced central-processing-unit runtime by approximately $4.1$-fold and remained applicable to the larger production fit. A Monte Carlo dropout neural comparator produced larger prediction errors and substantially wider predictive intervals. Six apparent thermokinetic structures were evaluated using mean-field variational inference, after which the nth-order and autocatalytic structures were validated using the No-U-Turn Sampler (NUTS). Under mean-field variational inference, the apparent autocatalytic structure achieved the lowest point estimate of the widely applicable information criterion (WAIC), the lowest derivative-domain error, and the lowest full-profile temperature-reconstruction root-mean-square error of $0.2920\mathrm{K}$. Its posterior obtained using NUTS yielded $E_a=40.98\mathrm{kJ}{\mathrm{mol}}^{-1}$, $k_{\mathrm{ref}}=0.005815{\min }^{-1}$, $\Delta T_{\mathrm{ad}}=56.11\mathrm{K}$, $m=0.1694$, and $n=1.0784$. The sampling diagnostics indicated satisfactory convergence, large effective sample sizes, and no divergent transitions. Although the MFVI posterior means and NUTS posterior medians were similar, variational inference produced narrower uncertainty intervals for several correlated parameters. Moving-block bootstrap intervals did not establish a decisive separation in WAIC among the leading structures. Expanded sensitivity, ablation, imperfect-insulation simulation, and endpoint-holdout analyses further showed that the apparent parameter estimates were sensitive to optimization, thermal nonideality, sensor response, and Gaussian-process boundary behavior. The autocatalytic formulation should therefore be interpreted as the best-performing apparent structure among the candidates tested rather than as evidence of a unique chemical mechanism. Overall, the framework extracted physically plausible apparent thermokinetic information from noisy temperature-only measurements while explicitly quantifying uncertainty arising from prediction, parameter estimation, model form, computation, thermal nonideality, and boundary behavior. Full article

Pipe welding automation requires accurate weld seam extraction and reliable robotic weld path generation under complex geometric conditions. Existing vision-based approaches often rely on expensive laser sensing systems, multi-view sensing, or continuous seam tracking, resulting in increased hardware cost and system complexity. To address these limitations, this study proposes a deep learning-based offline robotic welding framework that generates a three-dimensional welding path from a single top-view Red–Green–Blue and Depth (RGB-D) image acquired prior to welding. The proposed framework integrates weld seam detection, semantic segmentation, morphology-based post-processing, RGB-D image alignment, coordinate transformation, and polynomial-based trajectory refinement into a unified pipeline for robotic weld path generation. A custom pipe welding dataset consisting of 1476 annotated images collected from representative industrial pipe materials with varying diameters was constructed to evaluate the proposed framework. The experimental results demonstrate that the proposed Region of Interest (ROI)-guided weld seam extraction pipeline improves the U-Net segmentation performance from 0.735 to 0.791 mean Intersection over Union (mIoU), while the detection model achieves a Recall of 0.988 and an mean Average Precision at an Intersection over Union threshold of 0.5 (${\mathrm{mAP}}_{50}$) of 0.995. Furthermore, polynomial-based trajectory refinement reduces the three-dimensional positional root mean square error (RMSE) to 0.333 mm, enabling continuous robotic welding over the entire visible weld seam without additional path modification. These results demonstrate that the proposed framework provides a practical and cost-effective solution for offline robotic weld seam extraction and weld path generation, while establishing a promising foundation for future extension toward online robotic welding through real-time weld seam tracking and adaptive trajectory correction. Full article

Electric motors consume over 45% of global electricity and are a primary source of unplanned industrial downtime. Real-time fault detection at scale faces severe constraints, including distributed topologies, intermittent connectivity, and strict energy budgets on battery-powered edge nodes. Existing hierarchical federated learning approaches address resource disparities across tiers but lack downward feature transfer. This prevents resource-constrained edge sensors from utilizing cloud-learned representations when local fault data is sparse. This paper proposes a hierarchical federated cyber-physical architecture featuring three online cross-layer transfer mechanisms: warm-start Convolutional Neural Network (CNN) weight extraction, Long Short-Term Memory (LSTM) embedding alignment, and adaptive teacher–student distillation. This work is best characterized as a hierarchical federated-learning protocol and edge-hardware feasibility study: it validates the communication protocol, cross-layer transfer mechanisms, and sensor-tier hardware budget end-to-end, using the Gym-Electric-Motor (GEM) simulator as a controlled, reproducible, and openly available substitute for physically instrumented motor faults, rather than as a validated physical motor-fault diagnosis system. The framework is evaluated on a ten-node low-power System-on-Chip (SoC) microcontroller, low-power single-board computer, and cloud computing platform test bench, using GEM-simulated operating trajectories as a controlled, reproducible proxy for non-IID motor fault conditions in five-class motor fault detection. Under this test bench, the framework achieves an over 3-fold convergence speedup and reduces the sensor–cloud accuracy gap by nearly 74% (from 10.1% down to 2.7%) at a 200× lower compute budget. It improves minority-class diagnostic reliability, with F1 scores improving by 45% on average, while achieving over 95% sensor accuracy on this test bench, and the proposed mechanism also limits Macro-F1 degradation under injected sensor noise (SNR = 10 dB) to 8.0%, versus up to 22.5% for independent per-tier training. From an embedded electronics implementation perspective, the system operates under a 3.3 ms latency and consumes only 2.1 mJ per inference on the low-power SoC hardware—outperforming prior edge PdM deployments that report 4–6 mJ per inference—indicating a feasible architecture for energy-constrained edge intelligence, pending validation on physically measured fault data. Full article

Accurate vehicle trajectory prediction is essential for autonomous driving safety. However, existing data-driven models often ignore kinematic constraints, causing lateral jitter and trajectory distortion, while purely kinematics-based models lack flexibility in complex interactions. To address this, this paper presents a multimodal trajectory prediction method combining kinematics with dynamic interaction features. Operating in the Frenet coordinate system, the proposed model extracts historical features via a Bidirectional Gated Recurrent Unit (Bi-GRU) and utilizes an Adaptive Social Gating Network (ASGN) with multi-head attention to filter irrelevant interaction noise. This paper introduces a Stochastic Gating Decoder for multimodal latent variable sampling, adaptively fusing kinematics and data-driven paths to capture driver intention uncertainty while maintaining kinematic consistency. The model is trained using a composite loss function (Focal Loss and Best-of-K) to mitigate dataset long-tail distribution and trajectory divergence. Experiments on the HighD dataset show the proposed model achieves a minADE of 0.425 m and a minFDE of 0.955 m, outperforming baselines and reducing Lat-ADE by 53.9% compared to Social-GAN. These results confirm the model generates smoother, kinematically interpretable trajectories with higher accuracy in long-tail lane-changing scenarios. Full article