Abstract: This paper proposes a Spatio-Temporal-Physical Collaborative Track Initiation Network (STPC-Net) for short-sequence radar track initiation under low SNR and high false alarm density. Discriminative information from four-frame measurements is decoupled into three heterogeneous feature streams: a Mamba selective state space model capturing temporal evolution, a geometry-edge-biased fully connected graph attention network extracting spatial topology, and a FiLM mechanism injecting kinematic priors into learnable trajectory prototypes. A cross-modal gated fusion module integrates the three streams for collaborative inference. Experiments on real radar data yield a true candidate detection rate of 98.6% and a false alarm rate of 1.7%, outperforming threshold-based methods by 7-16 percentage points and deep learning baselines by approximately 2 percentage points. Under various measurement noise levels and true-to-false ratios of 1: 2-1: 8, detection degradation remains within 5 percentage points. Ablation studies confirm the complementary gains from heterogeneous feature synergy. The method provides high-quality initial tracks for downstream data association and track maintenance.