Bayesian Intent Prediction in Object Tracking Using Bridging Distributions.

Area of Science:

• Robotics and Artificial Intelligence
• Probabilistic Modeling
• Computer Vision

Background:

• Accurate intent prediction for tracked objects is crucial for effective decision-making in human-computer interaction, surveillance, and defense.
• Existing methods often struggle with predicting long-term object trajectories and destinations in advance.

Purpose of the Study:

• To develop a probabilistic inference approach for predicting the intended destination and future trajectory of a tracked object.
• To model object trajectories as Markov bridges to capture long-term dependencies and infer intent.

Main Methods:

• The proposed method models observed partial tracks as Markov bridges that terminate at the object's destination.
• Probabilities of possible destinations are evaluated by assessing the likelihood of partial tracks belonging to constructed bridges.
• The framework refines estimates of latent system states, predicts future values, and estimates time of arrival.

Main Results:

• A low-complexity Kalman-filter-based implementation is achieved for the inference routine.
• The approach effectively predicts object destinations and trajectories using both real-world (instrumented vehicle gestures) and synthetic (vessel navigation) data.
• Demonstrated effectiveness in predicting future states and arrival times.

Conclusions:

• The Markov bridge framework provides a robust method for advance prediction of object intent and trajectory.
• The probabilistic approach enhances system capabilities in applications requiring proactive decision-making based on object movement.
• The method offers a computationally efficient and versatile solution for trajectory prediction problems.