Railway foreign object intrusion detection is essential for intelligent railway monitoring systems. However, accurate real-time detection remains challenging due to small and distant targets, complex backgrounds, scale variations, adverse imaging conditions, and limited computational resources. To address these issues, this paper proposes a lightweight positional vision detection transformer, termed LPV-DETR, based on RT-DETR-R18. Specifically, a CSP-LGLB module is introduced into the backbone to enhance local–global feature representation, an AIFI-P module with learnable positional encoding is designed to improve spatial modelling, and a lightweight GVCFM module is developed for efficient cross-scale feature fusion. Experimental results on a self-constructed railway dataset show that LPV-DETR improves mAP@0.5 by 6.1 percentage points while reducing parameters and FLOPs by 27.1
Human pose estimation plays a fundamental role in understanding human behaviors and interactions in complex scenes. Among existing approaches, High-Resolution Networks (HRNet) and their variants have demonstrated strong performance in dense scenarios by maintaining high-resolution representations and enabling multi-scale feature fusion. However, the commonly used fusion strategies in HRNet are relatively simple, which limits their ability to effectively exploit cross-scale complementary information and thus affects robustness in challenging environments. To address these limitations, we propose C-HigherHRNet, an enhanced architecture built upon HigherHRNet. Specifically, a cross-scale dynamic fusion module is designed to adaptively aggregate informative features while suppressing redundant responses, improving the effectiveness of multi-scale representation learning. In addition, DySample-based upsampling is employed to enhance the quality of reconstructed feature maps, and skip connections are introduced to mitigate the loss of spatial details during multi-stage refinement. Extensive experiments on benchmark datasets, including COCO and CrowdPose, as well as a self-collected real-world dataset, demonstrate that the proposed method consistently improves keypoint localization accuracy. These results validate the effectiveness and generalization capability of C-HigherHRNet in handling complex scenarios with dense crowds and large-scale variations.
Infrared small target detection (IRSTD) is a critical task for airborne early warning and maritime surveillance, where weak targets with ultra-low pixel occupancy are easily submerged by structured clutter, making the balance between detection performance and false alarm suppression a long-standing challenge. Existing encoder–decoder methods face two key bottlenecks: (i) insufficient oriented long-range context modeling, leading to large-scale directional clutter being misidentified as target responses; (ii) unreliable boundary semantics in skip connections, which amplifies shallow noisy textures and spurious edges. To address these issues, we propose GlowNet, a Long-Range Oriented Network for single-frame IRSTD with lightweight design. Specifically, a Long-Range Oriented Context (LRoC) block is introduced at the bottleneck, which outperforms conventional axial convolution schemes by efficiently aggregating long-range oriented context to accurately distinguish directional clutter from targets. Meanwhile, an Edge-aware Gated Skip (EDGE-Gate) module is proposed, which surpasses typical edge gating mechanisms by selectively transmitting boundary-aware information while filtering noisy shallow features. Experiments on IRSTD-1K, SIRST-v1 and supplementary cross-dataset tests show that GlowNet achieves a superior detection-false alarm trade-off over the re-implemented MSHNet baseline under consistent evaluation protocols, with both pixel-level and target-level evaluation metrics. Meanwhile, GlowNet maintains ultra-low computation overhead, serving as a plug-and-play module with minimal additional complexity, which can be easily integrated into existing encoder–decoder IRSTD frameworks.
Multimodal large language models (MLLMs) can provide useful frame-level descriptions for video action recognition, but their outputs may be inaccurate or vary in relevance across a video. We propose CoKe-CLIP, a knowledge-enhanced framework that separately encodes category labels and MLLM-generated frame descriptions, extracts complementary RGB and edge representations, and uses edge–knowledge correspondence to compute temporal weights for aggregating frame descriptions. The aggregated knowledge representation is aligned with the category embedding as auxiliary training supervision. Experiments on Charades, HMDB51, and UCF101 show competitive closed-set recognition performance. We further examine category-response variation with a scale-controlled TSDS analysis; this analysis clarifies that the unnormalized TSDS decrease reflects the effect of temporal weighting on response scale and should not be interpreted as scale-independent semantic correction.
