Unmanned aerial vehicles (UAVs) equipped with radio-frequency (RF) sensing payloads provide flexible platforms for ground-emitter localization, but airborne sensing is often degraded by short dwell time, strong cochannel interference, aperture perturbation, navigation-state uncertainty, and intermittent RF-chain outages. These effects distort local direction-of-arrival (DOA) evidence and can lead to large localization errors after multi-UAV fusion. This paper proposes a localization-aware topology-adaptive covariance evidence integration framework, termed LA-TACR, for robust UAV-borne RF emitter localization. LA-TACR integrates constructive covariance reconstruction, reliability-tempered physical- and virtual-array bearing evidence, geometry-aware candidate selection, and robust position refinement within a MAP-motivated framework. Controlled Monte Carlo simulations show that LA-TACR consistently achieves lower DOA and localization errors than the compared practical RF estimators under few snapshots, vibration, element outages, navigation-state uncertainty, and coherent or intermittent interference. Additional sensitivity and angular-cue perturbation experiments demonstrate stable performance over the tested parameter range and gradual degradation under biased or partially unavailable angular information. Semi-measured experiments using the public DICHASUS-005x CSI dataset assess covariance and DOA evidence processing with measured channel backgrounds and controlled signal injection.
Accurate peak wavelength estimation from highly overlapped fiber Bragg grating (FBG) spectra is essential for high-density wavelength-division multiplexing. Although convolutional neural network (CNN)-based methods are promising, their performance strongly depends on the training dataset. Conventional datasets commonly use a Gaussian spectral model and random sampling. The Gaussian model does not accurately reproduce measured sidelobes, whereas random sampling rarely generates severe overlap cases. Severe overlap is defined as a four-peak wavelength span of 0.2 nm or less. Monte Carlo simulation showed that such cases account for only ~1% of a dataset generated by random sampling. To address both limitations, the proposed framework combines a Sinc²-Gaussian spectral model with random & overlap-oriented sampling. Spectral modeling and data sampling are examined systematically as explicit factors in CNN training dataset design. A CNN trained with the proposed dataset is experimentally evaluated using four FBGs sharing the same center wavelength. The method maintained high estimation accuracy over a strain range of 0–1000 με. Under severe overlap, the average coefficient of determination increased from 0.9749 for the conventional Gaussian model with random sampling to 0.9942 for the proposed dataset design. The maximum wavelength deviation also decreased from 49.4 to 17.5 pm. Under these conditions, the average standard deviation was ~2.8 pm. These results confirm that systematic training dataset design is critical for CNN-based demodulation of highly overlapped FBG spectra.
Autonomous underwater vehicle (AUV) docking requires reliable terminal pose sensing under optical degradation caused by turbidity, attenuation, and backscatter. In many docking stations, a fixed station-side camera is easier to power, calibrate, and maintain than an additional camera payload on a small AUV. However, station-mounted monocular sensing remains difficult because the approaching AUV appearance changes with range, illumination, and water clarity. This paper proposes RobustPercept, a turbidity-aware station-mounted monocular pose sensing framework for AUV docking. RobustPercept estimates an image degradation state from each camera frame and uses it as a learned reliability cue to modulate multi-scale features. During training, the predicted pose projects calibrated AUV keypoints into the image and is constrained by independently observed 2D keypoints through a visibility- and confidence-weighted robust reprojection loss. A conditioning analysis of the projection Jacobian exposes low-observability cases instead of treating all keypoint configurations as equally informative. Controlled simulation experiments show that RobustPercept achieves 93.4% perception success and 4.2 cm position error across rendered attenuation coefficients from β = 0.1 to 1.5 m−1. Water-tank validation using station-side camera images under four measured turbidity conditions further achieves 89.7% perception success and 5.3 cm frame-level mean position error, outperforming direct regression and detection-plus-PnP baselines. Jetson Orin NX profiling shows 14.8 ms FP32 and 6.5 ms INT8 network-inference latency on controlled terminal-docking imagery. These perception-only experiments do not evaluate closed-loop guidance, physical capture, collision avoidance, or recovery safety.
Vehicle positioning built around the Global Navigation Satellite System (GNSS) is insufficiently robust in road environments, whereas positioning methods based on on-board sensors are limited by a short observational range and blind spots. Considering the large-scale deployment of roadside surveillance with clear imaging and broad views, we propose a vehicle positioning method via monocular roadside imagery. Specifically, we effectively combine detected vehicles from roadside imagery with sparse geometry from vehicle models to establish reliable 2D-3D keypoint correspondences, which are then fed into a roadside camera-based pose solver to initially estimate the vehicle pose (position and orientation). Meanwhile, uncertainty modeling is conducted to quantify the propagation of keypoint prediction errors into the real road scene, and the modeled uncertainty is introduced into the pose solver to refine the pose estimate. Furthermore, an adaptive optimization strategy with keypoint uncertainty is employed to improve the pose estimate through appropriate iterations. Experimental results on the roadside surveillance dataset demonstrate that the proposed method achieves competitive positioning performance, with a mean positioning error below 0.22m and a 95% positioning error below 0.45m. Within the roadside surveillance area, it can accurately localize most vehicles while maintaining their crucial sparse 3D geometry.
