Mobile robots operating in structured indoor environments face significant challenges including the “kidnapped robot” problem, sensor accumulation errors, and environmental perceptual ambiguity, which collectively lead to slow convergence and inadequate accuracy in relocalization. To address these critical issues, this paper proposes an Adaptive Fusion Relocalization Algorithm (AFRA) based on a likelihood-field measurement model. The core innovations of AFRA include the construction of a refined likelihood-field model that effectively integrates LiDAR and Ultra-Wideband (UWB) data, significantly enhancing the accuracy of observation likelihood through probabilistic modeling of hybrid noise. Furthermore, a particle filter framework incorporating dynamic particle scheduling and adaptive resampling mechanisms is developed to achieve an optimal balance between precision and computational efficiency. The Experimental results demonstrate that AFRA maintains relocalization errors within ±0.035 m, improving accuracy by 45.3% compared to the best-performing single sensor, while achieving a 40.7% acceleration in convergence speed. These advancements substantially enhance the robustness and real-time performance of mobile robot localization in complex scenarios.