A Lightweight End-to-End Multi-Source Localization Framework in Urban Environments via Local Radio Map Construction

Abstract

Multi-source localization (MSL) plays a vital role in cognitive radio (CR) networks and spectrum monitoring by enabling timely and reliable positioning. Existing MSL methods typically follow a decoupled two-stage pipeline: they either regress source coordinates directly after estimating the source count, or first construct a spatial representation and then apply non-differentiable post-processing operations for coordinate estimation. However, such a disjoint design blocks gradient propagation and prevents end-to-end (E2E) optimization, leading to misaligned objectives and error accumulation, which ultimately limits localization performance. To overcome this, we propose a lightweight End-to-End MSL (E2E-MSL) framework that unifies local radio map construction with a novel differentiable localization module (DLM), enabling seamless gradient propagation and E2E optimization. The DLM, through differentiable coordinate selection (DCS) and physics-informed mean-shift (PIMS) clustering, allows the entire framework to be optimized in an E2E manner under a unified localization objective and naturally accommodates a variable number of sources without requiring architectural modifications or retraining. Extensive experiments on the VaryTxLoc dataset demonstrate that E2E-MSL significantly outperforms existing state-of-the-art (SOTA) approaches in accuracy, robustness, and inference efficiency, underscoring the critical advantage of E2E optimization for enhancing MSL performance. Our code is available at https://github.com/QLMSL/E2E-MSL