While machine learning (ML) has been explored for beam prediction, many methods remain constrained by single-modality inputs, shallow temporal modeling, and limited robustness to interference and domain shift. We introduce Multimodal Semantic-Enhanced Real-World Beam Prediction via Temporal Modeling with Visual Foundation Models (MSET), a multimodal framework that couples visual semantics with positional priors in a causal, lightweight design. A visual foundation model (VFM), instantiated as a Swin-Transformer, learns rich spatial and semantic cues from RGB images and Segment Anything Model (SAM)-derived region masks, and a lightweight ResNet-18 student distills this knowledge for efficient inference. Short-horizon dynamics are captured by a causal Temporal Convolutional Network (TCN) with an adaptive receptive field, whose volatility-driven depth gate expands context under motion spikes and contracts it in calm periods. On top of semantic-aware frame embeddings, a Temporally Aware Cross-Attention (TACA) aligns original and semantic-enhanced tokens, while a Mambaconditioned GPS prior, implemented via a selective state-space model (SSM), issues a location-conditioned single query with positional bias to attend the fused tokens. We further extend inference to dense deployments with multiple proximate candidates and address target selection under ambiguity. Experiments on DeepSense 6G dataset show consistent Top-k improvements across single/multi-target and day/night scenarios, indicating reduced sweep reliance and strong generalization under realistic V2I dynamics.

Unmanned Aerial Vehicles (UAVs) are crucial for wireless network transmissions, particularly in challenging en-vironments of regular inspection. However, transmitting high-resolution video data from UAV s poses challenges due to limited resources and significant data volumes. Traditional video compression methods, removing redundant information with a single frame, suffer from quality loss as compression rates increase. To address these issues, we propose a novel framework, namely Semantic-based Motion Detection Compression (SMDC) to perform the video compression with high-quality resolution. The proposed framework incorporates Generative Diffusion Change Detection (TransC-GD-CD), a robust semantic-based change detection method, to accurately detect motion between adjacent video frames. Specifically, frames with slight motion are eliminated, thereby reducing network bandwidth requirements for the UAV inspection. Furthermore, a neural network-based interpolation technique is integrated to restore the information loss and ensure smooth playback. Experimental results show that SMDC outperforms traditional compression methods based on H.264, achieving higher video quality at matched bitrates. The exceptional performance of SMDC promises its potential as an effective solution for high-resolution video transmission in scenarios with limited bandwidth.

Object detection (OD) in unmanned aerial vehicle (UAV) images faces many challenges, with diverse-scale objects and small objects being particularly prominent issues. To alleviate these challenges, we propose a novel multiscale feature learning and feature fusion network under the guidance of deformable convolution. First, a deformable convolution-guided feature learning (DCGFL) block is designed in the backbone to extract more effective multiscale features. The DCGFL block leverages the adaptability of deformable convolution to the shapes and scales of objects, akin to spatial attention. Moreover, it also employs channel attention to identify important feature maps. Hence, the proposed backbone possesses the functionality of spatial attention and channel attention. Second, in the neck, we devise a simple generalized feature pyramid network (SimpleGFPN) with several deformable convolution-guided feature fusion (DCGFF) blocks to fuse multiscale features. The proposed neck has cross-layer and cross-scale pathways, facilitating effective information exchange and fusion between shallow spatial and deep semantic features. Third, the SIoU loss is used to better model the bounding box regression loss. Finally, experimental results on the VisDrone2021 and UAVDT datasets show that the proposed method outperforms the compared OD methods. In terms of mean average precision, we obtain 37.8% on VisDrone2021 and 18.5% on UAVDT.