The detection and recognition of dim and small infrared (IR) targets across domains pose two formidable challenges: distributional discrepancies in samples and scarcity or absence of annotated instances in the target domain. While current unsupervised domain adaptive object detection methods can somewhat alleviate the performance degradation caused by these issues, they fail to address the differences in semantic content between the background environments in different application scenarios. This results in a semantic gap that impedes the algorithm’s adaptability. We propose a content-decoupled unsupervised domain adaptive method to mitigate the adverse impact of the semantic gap on domain adaptation. Specifically, we introduce an unsupervised task-guided branch in the one-stage detector that executes style transfer in real-time during training, directing the feature representation via shared parameters. A novel content-decoupled feature alignment module aligns semantically unrelated features in the source and target domains while preserving the intrinsic semantics of dim and small IR targets. The experiments on a cross-band IR dataset reveal that the proposed method enhances the mean average precision by 18.3% compared to the baseline detector without increasing computational burden.

Detecting infrared small targets quickly and accurately in complex backgrounds has always been a challenging task. Data-driven methods have achieved good results because of their powerful feature extraction capabilities. Many algorithms use ResNet or VGG as their backbone, but because of the small size and inconspicuous features, pooling layers in their networks could lead to the loss of targets in deep layers. Even though dense network structure is proposed to alleviate this issue, its excessive dense connections makes it difficult to achieve real-time detection. To meet the requirements of both accurate performance and real-time detection, we propose MDAFNet. We design a Moderately Dense Adaptive Feature Fusion (MDAF) module that contains only three feature layers as the backbone of the network. This module connects all internal features with each other and utilizes a weighted sum of different layers as the output, promoting feature reuse and maintaining infrared small target features in the deep layers of the network. We also design a Coarse-to-Fine detection head and introduce auxiliary loss to enable the network to predict target contours with greater precision. Moreover, we propose a new data augmentation method that effectively enhances the generalization performance of network. Experimental results demonstrate that our network achieves excellent performance in detection accuracy and meets the requirements for real-time detection on RTX3080 gpu.

In public safety scenarios, such as large-scale event security and urban crowd management, unmanned aerial vehicles (UAVs) serve as a vital tool for crowd localization, offering high mobility and broad coverage. However, UAV-based overhead localization faces challenges, including significant target scale variations due to altitude changes and poor feature visibility in low-light conditions. To overcome these issues, this study enhances the UAV-Dot framework by introducing a scale prediction branch for adaptive Gaussian heatmap adjustment, embedding a CBAM attention module in the U-Net encoder to strengthen feature extraction in dim environments and optimizing post-processing via dynamic thresholding and DBSCAN clustering. Experiments on the DroneCrowd dataset show that the improved model increases parameters by only 0.36% during training and 0.29% during testing yet achieves 53.38% L-mAP—outperforming the original UAV-Dot by 2.38% and STNNet by 12.93%. The model also delivers consistent gains of approximately 2% in L-AP@10, L-AP@15, and L-AP@20.

Infrared object detection in airborne surveillance faces significant challenges, particularly when targets are occluded and viewpoints are sparse and dynamic. Existing multi-view methods typically rely on dense camera arrays or precise static calibration, rendering them ill-suited for ad-hoc dual-UAV deployments characterized by asynchronous capture and sensor noise. To bridge this gap, we propose a Dual-Platform Collaborative Occlusion-Resistant Framework tailored for resource-constrained airborne platforms. This system-level architecture integrates three synergistic components to address these constraints: (1) a Hybrid Geometric-Semantic Registration (HGSR) scheme that establishes a robust spatial foundation by combining coarse physical projection with fine-grained semantic correction to align asynchronous views; (2) a Multi-Scale Enhanced Dual-Domain Strip Attention Module (MS-DSAM) that extracts occlusion-invariant features by jointly modeling spatial strip patterns and frequency-domain spectral consistency; and (3) a Dual-View Localization Loss (DLL), formulated as a weakly supervised regularization mechanism, which enforces cross-view semantic and geometric consistency without requiring dense supervision. Extensive experiments on the self-constructed DualView-Occlusion dataset and standard benchmarks (Wild-track, MultiviewX) demonstrate that our method achieves 58.9% mAP50 on small targets under moderate occlusion (Dataset A) using only two mobile views, outperforming state-of-the-art methods that rely on three or more fixed views. The system operates at 23.31 FPS, meeting real-time requirements for UAV edge deployment.