Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces a unified evaluation framework to compare deep-learning and classical DoA estimators for real-time ISAC beamforming, considering both accuracy and computational latency. By integrating various estimators with SVD-based beamforming and analyzing spectral efficiency, the study reveals distinct trade-offs in the accuracy-latency-performance design space. The Regression Convolutional Neural Network (RCNN) emerges as a promising candidate due to its favorable balance, while highlighting the need for beamforming strategies that mitigate SVD's computational burden for future low-latency ISAC systems.
Deep learning DoA estimators can drastically cut inference time, but SVD-based beamforming remains the primary latency bottleneck in real-time ISAC.
Integrated Sensing and Communication (ISAC) systems rely on accurate Direction-of-Arrival (DoA) estimation for effective beamforming and spectral-efficiency enhancement. While numerous classical and deep-learning-based DoA estimation methods have been reported in the literature, their suitability for real-time ISAC operation remains insufficiently understood because practical deployment requires consideration of both estimation accuracy and computational latency. This paper presents a unified evaluation framework for systematically comparing classical and learning-based DoA estimators within an ISAC beamforming pipeline. Representative classical and learning-based DoA estimators, together with an ensemble architecture, are integrated with Singular Value Decomposition (SVD)-based beamforming and evaluated under identical operating conditions. The impact of DoA estimation accuracy, inference latency, beamforming-weight generation time, and resulting effective spectral efficiency is investigated. Results show that different estimators occupy distinct operating points in the accuracy-latency-performance design space. Although learning-based approaches substantially reduce estimation time compared with MUSIC, the overall processing latency increasingly becomes dominated by the computational cost of SVD-based beamforming. Among the evaluated methods, the Regression Convolutional Neural Network (RCNN) exhibits a favorable balance between estimation accuracy, computational efficiency, and beamforming performance, making it a promising candidate for real-time ISAC deployment. The findings further suggest that future low-latency ISAC systems may benefit from beamforming strategies that reduce or avoid computationally intensive SVD-based beamforming-weight generation.