Integrated Sensing and Communication for Crossroads: AI-Driven End-to-End Target Classification and Range-Velocity Estimation Using Ray-Tracing Data

Abstract

Integrated Sensing and Communication (ISAC) is a key enabler for 6G systems, supporting urban crossroad sensing for intelligent transportation systems (ITS). However, reliable sensing of pedestrians and vehicles in such environments remains highly challenging due to severe static clutter and complex multipath propagation. Conventional ISAC frameworks operating in the delay-Doppler domain may not adequately preserve the fine-grained amplitude and phase information inherent in the original signal, thus limiting their sensing accuracy and robustness. In this paper, we propose an AI-driven end-to-end multi-task ISAC framework for urban crossroad scenarios. To support this framework, we construct a physics-aware OFDM ISAC dataset using the NVIDIA Sionna ray-tracing engine, which explicitly captures the distinct electromagnetic characteristics of pedestrians and vehicles. We design a multi-task ISAC learning model that jointly performs target classification, range estimation, and velocity estimation directly from the channel frequency response (CFR) matrix, without relying on delay-Doppler processing. The proposed model consists of a shared feature extraction backbone and task-specific attention modules. An uncertainty-weighted loss function is employed to balance multi-task learning, and an SNR-driven curriculum learning strategy is introduced to ensure robust convergence. Extensive evaluations demonstrate that by directly leveraging the raw CFR, our proposed framework can preserve critical phase signatures and significantly outperform conventional delay-Doppler baselines in multi-task sensing under severe static clutter.