Cross-Camera Pedestrian Trajectory Retrieval Based on Linear Trajectory Manifolds.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Pedestrian trajectory retrieval is vital for analyzing crowd flow and identifying individuals across camera networks.
• Traditional methods require extensive spatiotemporal data and camera positional information, posing significant data collection challenges.
• Existing approaches often struggle with the complexity of multi-camera pedestrian tracking.

Purpose of the Study:

• To develop a novel method for pedestrian trajectory retrieval that bypasses the need for spatiotemporal modeling.
• To introduce an Implicit Trajectory Encoding scheme, Temporal Rotary Position Embedding (T-RoPE), for enhanced feature representation.
• To improve the efficiency and accuracy of multi-camera pedestrian tracking.

Main Methods:

• Proposed a Temporal Rotary Position Embedding (T-RoPE) scheme to encode temporal information directly into visual representations.
• Developed a method to model inter-camera trajectory extraction within a refined feature space using a linear trajectory manifold.
• Utilized visual characteristics of candidate trajectories for comparison and ranking against query features.

Main Results:

• The T-RoPE module effectively integrates temporal data, creating a novel feature space for trajectory analysis.
• The proposed method successfully addresses inter-camera trajectory extraction challenges by identifying linear trajectory manifolds.
• Experiments demonstrated significant improvements in pedestrian trajectory retrieval precision across diverse datasets.

Conclusions:

• Temporal Rotary Position Embedding (T-RoPE) offers a versatile, plug-and-play solution for enhancing pedestrian trajectory retrieval.
• The method reduces reliance on complex spatiotemporal data, simplifying the retrieval process.
• The newly introduced Mall Trajectory Dataset and released code facilitate further research in this area.