Abstract
Real-time obstacle avoidance and spectrum sensing in complex environments conflict inherently, making accurate radiation source tracking difficult. This letter establishes an online learning framework by proposing a progressive fusion algorithm combining variational inference and Gaussian processes. An end-to-end neural network that achieves linear-complexity inference with $O(HW + M + nm)$ complexity per time step by mapping depth images and spectrum data into time-optimal trajectories is the main innovation. Through trajectory primitive coordination with REM construction, a dual-update mechanism synchronously optimizes localization and obstacle avoidance. Using Gaussian processes to model positional uncertainty, the variational Bayesian inference engine enables probabilistic source prediction without specialized positioning modules. Simulations show 10-second exploration-localization completion in $30\times 60$ m2 areas and collision-free navigation at 4 m/s in obstacles spaced 0.3-0.6m. Comprehensive high-fidelity simulation studies validate the algorithm’s robustness across diverse environments and radiation source configurations. Compared to traditional optimization-based methods with $O(N^{3})$ complexity and particle filter approaches, our algorithm achieves 15% better localization accuracy, 25% faster convergence, and 40% reduced computational cost while maintaining real-time performance with 15 ms per control cycle. The fusion of model-driven REM and data-driven planning exhibits superior adaptability over conventional spectrum mapping approaches.

This publication has 10 references indexed in Scilit: