11 December 2025 in Pattern Recognition
23 October 2026 in Pattern Recognition
4 October 2026 in SSRN Electronic Journal
Front-view RGB--LiDAR semantic segmentation requires the joint exploitation of dense visual appearance and sparse geometric observations within a unified image plane. However, projected LiDAR is characterized by large empty regions, discontinuous local responses, and heterogeneous geometric-variable distributions, making standard patch embedding designed for dense image signals suboptimal for such non-uniform sparse inputs. To address this issue, we propose a sparse-aware geometric tokenization framework for front-view RGB--LiDAR semantic segmentation. Specifically, a Sparse-Aware Geometric Embedding (SAGE) module organizes three-dimensional Cartesian coordinates together with deterministically derived range and angular variables into a six-channel geometric representation, and integrates local response aggregation, token normalization, and empty-patch initial-response suppression into the LiDAR token construction process. Within a dual-stream Transformer architecture, modality-specific hierarchical reconstruction is further employed to maintain independent token reassembly and spatial recovery paths for RGB and LiDAR, thereby preserving modality-specific token-to-pixel propagation before cross-modal fusion. A lightweight Cross-modal Locality-Preserving Refinement (CLPR) module is additionally introduced before scale-wise fusion to refine local LiDAR representations using aligned RGB context. Experimental results show that introducing the hierarchical reconstruction framework improves Vehicle IoU and Human IoU from 91.33\% and 66.91\% to 94.03\% and 76.38\%, respectively. Incorporating SAGE further increases the two metrics to 95.07\% and 79.10\%, while the complete model achieves 95.09\% Vehicle IoU and 79.61\% Human IoU. Additional analyses demonstrate that SAGE consistently yields positive gains across different LiDAR geometric-support levels and road-recording conditions, whereas the contribution of CLPR is mainly concentrated on a subset of challenging Human boundary regions. These results indicate that sparse-aware geometric tokenization can effectively complement hierarchical dual-stream spatial reconstruction and improve pixel-level RGB--LiDAR semantic segmentation in front-view driving scenes. The code will be released upon acceptance.
29 May 2026 in SSRN Electronic Journal
Mobile Edge Computing (MEC) empowered by Unmanned Aerial Vehicles (UAVs) has emerged as an effective approach for handling the substantial data-processing tasks generated by widely distributed User Devices (UDs). By exploiting flexible mobility and favorable air-to-ground communication links, UAVs can act as aerial edge servers to improve task processing efficiency. However, practical deployments are constrained by the need to properly schedule UAV movements and by limited onboard computing resources, which together may lead to excessive latency for real-time services. In this paper, we study a UAV-assisted MEC system with the objective of minimizing overall delay, where the delay metric incorporates both queueing and computation delays under an energy constraint. We jointly optimize UAV flight trajectories and computation offloading decisions to reduce service time. Due to the strong coupling between continuous trajectory variables and offloading decisions, the resulting problem is difficult to solve using standard convex optimization techniques. To obtain a scalable solution, a joint design is formulated as a Markov Decision Process (MDP) and a cooperative multi-agent deep reinforcement learning framework is provided. A Twin Delayed Deep Deterministic (TD3) policy gradient scheme is employed to better handle large-scale continuous state–action spaces and to enhance learning stability. Simulation results indicate that the proposed approach attains a reduced average latency compared to the deep deterministic policy gradient baseline.
8 June 2026 in SSRN Electronic Journal
Monocular 3D visual grounding aims to localize the object referred to by a natural language expression in a single RGB image and estimate its 3D attributes. Despite recent progress, existing methods usually rely on globally coupled vision-language interactions and often lack an explicit transition from 2D semantic grounding to 3D geometric reasoning, which limits robustness in cluttered, ambiguous, and long-range scenes. To address this issue, we propose a Text-guided Depth-aware Monocular 3D (TDM3D) visual grounding framework. The key idea is to decompose the task into a structured grounding process, where reliable vision-language correspondence is established before 3D spatial reasoning is performed. Specifically, a Region-Aware Enhancement Head (RAEH) is introduced to strengthen object-centric visual responses and suppress background interference, while depth-aware embeddings are used to inject geometric priors for 3D reasoning. Furthermore, a two-stage decoder is designed to progressively bridge 2D semantic grounding and 3D geometric modeling. Experiments on the Mono3DRefer benchmark demonstrate that the proposed method achieves 53.14%, 59.07%, and 57.95% Acc@0.5 on the Unique, Multiple, and Overall settings, respectively, outperforming existing state-of-the-art methods. Extensive ablation studies further verify the effectiveness of the progressive decoding strategy and the region-aware enhancement design. These results show that explicitly modeling the progression from semantic grounding to geometric reasoning provides a more effective solution for monocular 3D visual grounding. The code will be released upon acceptance.