Algorithm 1. Screened IMM-GRU filtering-tracking algorithm

Input: Radar measurement sequence $Z_k$, sequence length L, four filtering networks ${\{f_i(\cdot )\}}_{i=0}^3$, classification network $C(\cdot )$, polar-coordinate noise covariance $R_k$, integer fixed-lag order n, and CA process-noise intensity $q_s$.

Output: Final position estimate ${\widehat{p}}_k^S$ obtained by screened filtering.

Step 1. For each time $k$, update the measurement sequence of length L and obtain the normalized measurement sequence ${\overline{Z}}_k$ by min-max normalization.

Step 2. Input the normalized measurements into the four GRU filtering networks to obtain ${\widehat{p}}_k^{(i)}$,$i=1,2,3,4$.

Step 3. Construct the convex feasible domain ${\mathcal{D}}_k$ of the current network from the outp-uts of the four filtering networks.

Step 4. Convert the polar measurement into a Cartesian measurement and obtain $R_k^c$ through Jacobian propagation.

Step 5. Perform forward Kalman filtering under the CA model to obtain $s_k^f,P_k^f,s_k^{-},P_k^{-}$.

Step 6. For each state time, take $m_k=\min (N,k+n)$ and perform RTS backward smo-othing over the interval $[k,m_k]$ to obtain the reference position $r_k$.

Step 7. Project $r_k$ onto the network convex feasible domain ${\mathcal{D}}_k$ to obtain ${\widehat{p}}_k^S={\Pi }_{{\mathcal{D}}_k}(r_k)$.