Potential Field-Trajectory Correlation Model (PTCM)

• The paper demonstrates that combining semantic fusion with a hybrid A* search cuts occlusion duration by over 70% and reduces tracking failure rates below 5%.
• PTCM is a unified framework that integrates intention-aware prediction, geometric reachable region construction, and constraint-based trajectory optimization for robust aerial tracking.
• The model’s real-time performance and empirical success in agile, safety-critical maneuvers highlight its transformative potential in immersive aerial target pursuit.

The Potential Field-Trajectory Correlation Model (PTCM) designates a unified framework for robust and anticipatory trajectory planning in aerial target tracking, as typified by the Intention-Driven Area Prediction Module (IDAPM) introduced for quadrotor-based target pursuit. This model integrates semantically informed intention prediction—drawing from both environmental context and low-level state observation—with dynamic geometric reachability analysis, resulting in refined trajectory generation that is sensitive to predicted future target maneuvers. The PTCM governs the interaction between intention-aware prediction, hybrid discrete-continuous motion planning, and constraint-based trajectory optimization, producing real-time, safety-critical, and visibility-preserving flight paths for agile tracking platforms (Ren et al., 2023).

1. Semantic Intention Fusion in Trajectory Prediction

The PTCM first formalizes a finite set of discrete intentions $I$, each representing a meaningful maneuver class (e.g., constant velocity $i_{cv}$, turn-left $i_{tl}$, turn-right $i_{tr}$, decelerate $i_{dec}$), with $m_i=4$ in the cited work. For each $i \in I$, the model associates two real-valued signals at each time $t$:

• Potential assessment $R_i(t)$: Encodes the likelihood of intention $i$ becoming optimal, based on environment/kinematics.
• State observation $O_i(t)$: Quantifies recent target motion evidence supporting $i$.

These are fused using

$$\Pi_i(t) = \frac{1}{2}\left(\tanh(R_i(t) + O_i(t) - b_0) + 1\right)$$

yielding a probability-like belief for each intention. Specific instantiations depend on geometric/kinematic cues: for turn intentions, $R_{i_{tl}}$ is a function of the difference in sector half-angles over a window $t_0$, and $O_{i_{tl}}$ depends on the angular divergence between body orientation and velocity. Deceleration probability increases as target speed rises and as obstacles approach.

This semantic fusion mechanism allows real-time, data-driven estimation of target intent, which serves as a prior for subsequent prediction modules (Ren et al., 2023).

2. Geometric Reachable Region Construction

Independently from semantic fusion, PTCM builds a time-varying sector-shaped reachable set, $$\mathcal{R}(t) = \mathcal{R}_l(t) \cup \mathcal{R}_r(t)$$, centered at the filtered target position $(P_t, V_t)$, projected by a constant-velocity model over a limited horizon $t_0$. Each half-sector $\mathcal{R}_{l/r}(t)$ is defined by the inner product relative to its bisector axis $\xi_{l/r}$ and corresponding half-angle $\theta_{l/r}$. The sector geometry adapts in real time to local obstacle clearance, permitting asymmetric "fan-out" on the less-constrained side, which directly informs intention potentials for anticipated turns.

This explicit, adaptive geometric model bounds predicted target motion while embedding environmental context—a core tenet of the PTCM design.

3. Intention-Driven Hybrid A* Search for State Trajectory Generation

PTCM leverages a modified 6D Hybrid A* search over $(p, v)$ state space, orchestrated via intention-specific motion primitives. For each node, the chosen intention not only governs the transition dynamics (constant-velocity for $i_{cv}$, coordinated turn for $i_{tl}/i_{tr}$, constant-acceleration for $i_{dec}$; see Table I in (Ren et al., 2023)) but also contributes to cost via a penalty matrix $\Lambda$ that discourages excessive intention switching. The objective integrates both an intention-switch penalty $g(n)$ and a spatial heuristic $h(n)$, guiding the search toward plausible and efficient future positions $\{z_k, t_k\}$ over the prediction horizon.

The output is a temporally and semantically annotated sequence of trajectory "gates"—localized prediction points, each tagged by the driving intention, for downstream trajectory optimization.

4. Area Prediction, Uncertainty Quantification, and Intention-Aware Constraints

To interface between prediction and control, each predicted state $z_k$ is enveloped in two convex regions:

• Occlusion-free flight corridor $\phi_k$ computed via Elastic Tracker algorithms.
• Visibility sector $V_k$ and intention-informed enlargement $\hat{V}_k$, with half-angles $\hat{\theta}_k$ expanded proportionally to the left/right turn probabilities ($\Pi_{i_{tl}}, \Pi_{i_{tr}}$).

Additionally, minimum safe distances are dynamically scaled according to the predicted deceleration probability ($\hat{d}_l = d_l + d_\epsilon \Pi_{i_{dec}}$), offering calibrated spatial buffers in high-uncertainty regimes.

The explicit uncertainty representation enables downstream optimizers to enforce robust, risk-aware constraints during trajectory synthesis, particularly when the target's future behavior is ambiguous or likely to change abruptly.

5. Integration with Spatial-Temporal Trajectory Optimization

Given the set of predicted points, corridors, and dynamically modulated constraints, the PTCM framework employs MINCO polynomial trajectory optimization. The cost functional $J(q, T)$ aggregates smoothness, time efficiency, penalty terms for violation of flight corridor boundaries, and absolute intention-aware constraints at each $t_k$:

• Enforcing $p(t_k) \in \hat{V}_k$ to guarantee FOV maintenance,
• Maintaining $p(t_k) \in V_{k+1}$ for additional visibility margin,
• Constraining $\hat{d}_l \leq d_{xy}(p(t_k), z_k) \leq d_u$ to ensure safe tracker-target separation.

Constraint violations are penalized cubically, and gradients are computed analytically for efficient backpropagation through MINCO's mapping. This formulation ensures continuous, feasible, and intention-consistent tracking trajectories, fully leveraging the probabilistic and geometric output of previous PTCM stages.

6. Parameterization, Assumptions, and Empirical Performance

PTCM exposes tunable parameters, including:

• $k_1,...,k_6$, $b_0,...,b_2$ (intention fusion/scaling),
• $t_0$ (anticipation window duration),
• $\theta_0$, $\theta_\epsilon$ (visibility margin angles),
• $d_\epsilon$ (dynamic distance margin),
• $w_h$ (A* heuristic weight), $\Lambda$ (intention-switch penalties),
• $T_p$ (prediction horizon), $\Delta t$ (primitive step), $N_p$ (number of predicted points).

The model relies on several operational assumptions:

• Target motion is well-approximated by CV/CT/CA primitives,
• Four intentions suffice for tracking granularity,
• Global obstacle information is accessible for region construction,
• EKF yields accurate real-time $(P_t, V_t)$ estimates.

Reported empirical results demonstrate that, at high target speeds (1.5 m/s), the IDAPM implementation achieves over 70% reduction in occlusion duration and decreases the tracking failure rate to below 5% compared to a 30% failure rate in intention-unaware schemes. In real flight benchmarks, success rates in T-junction navigation improved to 80% from 30%, with planning operating at over 25 Hz and sub-100 ms latencies on standard computing hardware (Ren et al., 2023).

7. Significance and Implications within Aerial Tracking

PTCM, as embodied by the IDAPM, represents a principled synthesis of semantic intention models, geometric reachability, discrete-continuous planning, and uncertainty-aware optimization, each grounded in operational constraints and real-time requirements. This approach directly addresses the limitations of conventional, purely kinematic predictors, showing substantial empirical gains in both safety and robustness under adversarial and ambiguous motion. A plausible implication is that such modular, semantically enriched models are extensible to broader domains requiring online prediction and planning amidst uncertainty, with structural flexibility for additional intentions, richer intention state spaces, or more complex environments (Ren et al., 2023).