Two-Stage Multi-Frame Fusion Model for Radar Anti-Jamming Target Detection with Range-Doppler Maps

Abstract

Mainlobe active deception jamming significantly degrades radar target detection in complex electromagnetic environments. Current anti-jamming detection approaches inadequately exploit the spatiotemporal information inherent in target echoes, thereby fundamentally limiting their capacity for reliable detection and robust true–false discrimination. Therefore, this paper proposes a trajectory-guided two-stage multi-frame fusion detection (TSMF-Det) model based on range-Doppler (R-D) maps. Initially, a lightweight candidate detector is utilized to generate high-recall candidate boxes, which are then linked into cross-frame trajectories via temporal association. Subsequently, a spatiotemporal refinement network (STRNet) is designed to aggregate multi-frame temporal and multi-scale spatial features, enabling fine-grained true–false target discrimination and bounding-box refinement. Finally, an adaptive peak-center fusion module dynamically achieves high-precision estimation of range and velocity parameters. Experiments on a comprehensive dataset encompassing six typical active deception jamming scenarios demonstrate that TSMF-Det yields an average detection probability of 0.97. Comparative experiments show that TSMF-Det maintains stable target detection under strong jamming while providing effective false-target rejection and accurate range–velocity estimation. By exploiting multi-frame spatiotemporal features, TSMF-Det provides an effective anti-jamming detection framework for cognitive radar systems.