Multi-target localization and trajectory prediction via single-pixel detection in complex backgrounds

High-frame-rate multi-target tracking remains challenging for conventional vision systems, particularly in complex scenes. This paper proposes an SPI-LSTM fusion framework that integrates trajectory prediction into single-pixel imaging (SPI). A lightweight long short-term memory (LSTM) network is employed to learn generalized kinematic information, while a dual-stage arbitration mechanism coordinates measurement and prediction by switching between Fourier-based localization and LSTM prediction during target crossovers. Benchmarking against a conventional CMOS-based vision system shows that the proposed approach achieves higher tracking accuracy while requiring less than 0.5% of the data bandwidth. For a 256×256 scene, simulations achieve localization and prediction RMSEs below 1.08 and 1.63 pixels. Laboratory experiments reach 49 fps with localization and prediction errors below 2.23 and 2.91 pixels, while outdoor daylight experiments achieve 108 fps with errors of 3.77 and 2.43 pixels. These results demonstrate that the proposed framework enables accurate and high-speed multitarget localization and prediction with extremely low data bandwidth, highlighting its potential for bandwidth-constrained sensing applications.