Joint Intent & Trajectory Prediction (MTITP GAN)

Fri, 01 May 2026 00:00:00 +0000

Joint Intent and Trajectory Prediction (MTITP) is the generative sequence modeling framework developed in Chapter 7 of my doctoral dissertation. Rather than treating intent recognition and trajectory forecasting as separate problems, MTITP couples them in a unified multi-task network.

Core Innovations & Architecture

MTITP-WGAN Framework: Formulates a Multi-Task Intent and Trajectory Prediction Generative Adversarial Network trained with Wasserstein loss and gradient penalty (WGAN-GP) to eliminate mode collapse and generate realistic, multi-modal future paths.
Three-Pronged Joint Output:
1. Past-Intent Classification: Recognizes behavioral intent over historical encounter windows.
2. Future-Intent Forecasting: Predicts forward-looking tactical intent transitions.
3. Intent-Conditioned Trajectory Synthesis: Generates kinematically feasible future coordinate sequences conditioned on predicted intent.
Robustness Under Sensor Noise: Evaluated extensively across simulator-generated benchmarks with varying noise regimes (Noiseless, Noisy 1, Noisy 2) and out-of-distribution adversarial encounters (herding, ramming, blocking).
Ablation & Baselines: Rigorously benchmarked against MarITGAN v2 and non-adversarial variants (MTITP-L2) to isolate the exact contribution of adversarial loss to trajectory fidelity.

Forthcoming Publications

Two pending journal manuscripts are currently derived from this framework:
- Joint Multi-Task Intent and Trajectory Prediction for Autonomous Maritime Surface Vessels using Wasserstein GANs (Target: IEEE Transactions on Intelligent Transportation Systems (T-ITS))
- Generative Scenario Augmentation and Counterfactual Intent Analysis in Safety-Critical Maritime Encounters

Maritime Trajectory Prediction & State Estimation

Wed, 15 Apr 2026 00:00:00 +0000

Maritime Trajectory Prediction & Bayesian State Estimation forms the foundational empirical benchmarking contribution of Md Abu Sayed’s doctoral dissertation (Chapter 6; TrajectoryKF).

Before deploying complex deep generative models, operational autonomy systems require establishing the exact performance boundary of classical Bayesian filters and kinematics estimators. This project provides a comprehensive, fine-grained empirical study comparing six canonical filter architectures across multi-vessel encounter dynamics.

Evaluated Filter Architectures

Constant Velocity (CV) Baseline: Linear state propagation under Gaussian motion assumption.
Standard Unscented Kalman Filter (UKF-4D): Nonlinear sigma-point propagation with standard 4D kinematic state [x, y, v_x, v_y].
UKF with 5D Constant-Velocity State (UKF-5D-CV): Augments the state with heading (θ) for coordinated forward extrapolation.
Asymmetric UKF with Acceleration State (UKF-CA-6D): Models constant acceleration [a_x, a_y] to capture rate changes during maneuvering.
Extended Kalman Filter with Constant Turn Rate & Velocity (EKF-CTRV): Analytic Jacobian linearization of circular arc kinematics [x, y, v, θ, ω].
UKF-CTRV: Deterministic sigma-point propagation through exact CTRV nonlinear dynamics.

Key Findings & Research Questions