Towards Trajectory Forecasting From Detection.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Accurate trajectory forecasting is crucial for autonomous systems.
• Existing methods rely on pre-extracted, often noisy, ground truth trajectories, leading to prediction errors.

Purpose of the Study:

• To develop a robust trajectory forecasting method that does not require explicitly formed trajectories.
• To improve the reliability of trajectory prediction in the presence of noisy detection and tracking data.

Main Methods:

• Proposes predicting trajectories directly from detection results, avoiding reliance on explicit trajectory extraction.
• Introduces an affinity-aware state update mechanism to manage state information based on detection affinities.
• Incorporates aggregation of states from multiple plausible matching candidates to handle association uncertainty.

Main Results:

• The proposed method demonstrates improved robustness against noisy trajectory data.
• Experimental validation confirms the effectiveness and generalization capabilities of the approach.
• The method successfully mitigates forecasting errors caused by data association inaccuracies.

Conclusions:

• Directly predicting trajectories from detections offers a more robust alternative to methods relying on ground truth trajectories.
• The affinity-aware state management and candidate aggregation effectively handle uncertainty in data association.
• This approach enhances the reliability of trajectory forecasting for autonomous platforms.