Unsupervised domain adaptation (UDA) aims to adapt a model trained on the source domain (e.g. synthetic data) to the target domain (e.g. real-world data) without requiring further annotations on the target domain. Most previous UDA methods for semantic segmentation focus on minimizing the domain discrepancy of various levels, e.g., pixels and features, for extracting domain-invariant knowledge. However, the primary domain knowledge, such as context and detail correlation, remains underexplored. To address this problem, we propose a context- and detail-enhanced unsupervised learning framework, called CDEA, for domain adaptive semantic segmentation that facilitates image detail correlations and contexts semantic consistency. Firstly, we propose an adaptive masked image consistency module to enhance UDA by learning spatial context relations of the target domain, which enforces the consistency between predictions and masked target images. Secondly, we propose a detail extraction module to enhance UDA by integrating the learning of spatial information into low-level layers, which fuses the low-level detail features with deep semantic features. Extensive experiments verify the effectiveness of the proposed method and demonstrate the superiority of our approach over state-of-the-art methods.

Autonomous aerial refueling (AAR) can improve combat radius and endurance, and the visual detection of drogue and airplane is a key technology especially for autonomous docking of unmanned aerial vehicles (UAVs). However, limited computing resources of airborne instruments, multi-scale and complex scenes seriously affect the performance and robustness of visual detection. To address the above challenges, we propose a real-time and multi-scale drogue and airplane visual detection method. Firstly, we design a lightweight backbone based on structural re-parameterization to reduce computational cost and improve inference speed. Secondly, the multi-scale feature enhancement is proposed to learn significant differences in feature scales between airplane and drogue by enabling global receptive field and encoding positional information. Finally, we design an efficient decoupled head to decouple classification and regression branches, and propose a regression loss function to avoid width-height gradient contradiction. Experimental results show that the detection accuracy of our method is superior to existing SOTA methods, and the inference speed of 128.50 FPS is achieved on embedded instruments, which verifies the robustness and real-time performance of the proposed method.

Robotic navigation tasks require multi-dimensional sensors collaboration for environmental perception, where data fusion between camera and lidar plays a critical role in 2D-3D data fusion. However, the reliability of current 2D-3D sensors fusion approaches is limited by the accuracy, efficiency, and working range of joint sensors calibration. To address the challenges posed by the complexity and wide-range requirements of 2D-3D sensors calibration, this paper analyzes the characteristics of 2D and 3D data and proposes a novel multi-feature calibration target. This target integrates both checkerboard patterns and spherical markers, levesraging the camera’s sensitivity to planar features and the lidar’s ability to perceive spatial structures, thus enabling high-precision feature extraction for both sensors modalities. In addition, spatial alignment between the measurement domains of the two sensors is performed to achieve high-precision calibration of structural parameters across the entire measurement range. Experimental results demonstrate that the proposed joint calibration method achieves high accuracy and provides a reliable calibration foundation for robotic navigation and other 2D-3D sensors fusion tasks.