CAMP: A Context-Aware, Multimodal, and Privacy-Preserving Pedestrian Trajectory Prediction Framework.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Pedestrian trajectory prediction is crucial for crowd analysis and human-robot interaction.
• Current deep learning models improve accuracy but lack transparency and privacy for individual motion patterns.

Purpose of the Study:

• To introduce CAMP, a Context-Aware, Multimodal, and Privacy-preserving framework for pedestrian trajectory prediction.
• To enhance prediction accuracy while mitigating privacy risks associated with individualized motion data.

Main Methods:

• Utilized a role-aligned multimodal architecture with separate encoders for trajectory, optical flow, and spatial interactions.
• Employed a Transformer-based decoder for fusing multimodal inputs to predict future trajectory distributions.
• Implemented targeted Differential Privacy by Stochastic Gradient Descent (DP-SGD) on the individual motion branch for privacy preservation.

Main Results:

• CAMP achieved competitive performance (ADE/FDE) on the ETH/UCY benchmark.
• The privacy-preserving variant, DP-CAMP, demonstrated a viable utility-privacy trade-off across various privacy budgets.
• The framework effectively separates shared motion patterns from individualized tendencies.

Conclusions:

• CAMP offers an effective and privacy-conscious approach to pedestrian trajectory prediction.
• The proposed method balances prediction accuracy with the critical need for data privacy.
• This work advances the development of transparent and secure AI systems for human-centric applications.