Multi-Camera Trajectory Forecasting With Trajectory Tensors.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Existing trajectory forecasting methods primarily focus on single-camera scenarios (SCTF).
• Single-camera approaches have limited field-of-view, hindering long-term trajectory prediction.
• Multi-camera systems are crucial for surveillance and traffic monitoring but lack robust forecasting tools.

Purpose of the Study:

• To address the limitations of SCTF by developing a multi-camera trajectory forecasting (MCTF) framework.
• To enable accurate prediction of object trajectories across a network of cameras.
• To introduce a novel approach for handling object appearances and locations across multiple viewpoints.

Main Methods:

• Developed a Which-When-Where framework for MCTF.
• Proposed trajectory tensors for encoding multi-view trajectories and uncertainties.
• Created and utilized a new database with 600 hours of video data from 15 camera views for MCTF task.

Main Results:

• Proposed trajectory tensor models demonstrated superior performance compared to coordinate trajectory-based MCTF models.
• The new MCTF framework outperformed adapted SCTF methods.
• Experiments validated the effectiveness of the trajectory tensor approach on a dedicated multi-camera dataset.

Conclusions:

• The trajectory tensor approach provides a robust solution for multi-camera trajectory forecasting.
• The developed MCTF framework enhances the capabilities of surveillance and traffic monitoring systems.
• This work advances the field by enabling more accurate and comprehensive object tracking in complex multi-camera environments.