Algorithm 2. Screened IMM-GRU prediction-tracking algorithm

Input: Prediction input sequence $B_k$, previous-time model probability ${\pi }_{k-1}$, four pred-iction networks ${\{g_i(\cdot )\}}_{i=0}^3$, classification network $C(\cdot )$, model transition matrix $\Pi$, and screening parameter set $\Omega$.

Output: The predicted position ${\widehat{p}}_{k+1}$ and predicted variance ${\widehat{\Sigma }}_{k+1}$ at the next time step.

Step 1. The classification network gives the current probability $q_k=C(B_k)$, and ${\pi }_{k-1}$, $\Pi$ is used to propagate the previous-time probability to obtain the fused probability ${\pi }_k$.

Step 2. Call the four prediction networks to obtain ${\mu }_{k+1}^i$ and $v_{k+1}^i$, $i=1,2,3,4$.

Step 3. For $B_k$, take its last three positions $b_{k-2}$$, b_{k-1}$, and $b_k$, calculate $D_i$$, S_i$$, \mathrm{and} A_i$, and further obtain the proxy uncertainty $u_i$.

Step 4. Sort the networks according to descending ${\pi }_k^i$, and when the cumulative probability reaches ${\tau }_{\mathrm{post}}$, construct the initial candidate network set ${\mathcal{R}}_k^0$; if the classification entropy $H_k$ is too large or the gap between the two largest probabilities $\Delta \pi$ is too small, additionally retain the maneuver network to obtain the final candidate network set ${\mathcal{R}}_k$.

Step 5. Compute the score $u_i$ within the candidate set ${\mathcal{R}}_k$, and select the $K_f$ networks with the smallest scores to form ${\mathcal{S}}_k$.

Step 6. Within ${\mathcal{S}}_k$, compute the normalized weights $w_i$, and complete mean fusion and variance moment matching.

Step 7. Write the fused mean, real flag 0, variance trace, and incremented age back to the buffer; update the model probability and record the prediction result.

Step 8. Output the predicted position and predicted variance.