[Paper Review] Distance-based Camera Network Topology Inference for Person Re-identification

Yeong-Jun Cho, Kuk‐Jin Yoon|arXiv (Cornell University)|Dec 1, 2017

Video Surveillance and Tracking Methods52 references3 citations

TL;DR

This paper proposes a distance-based camera network topology inference method for person re-identification that improves accuracy in large-scale surveillance by estimating inter-camera distances using calibrated camera parameters and human walking speeds. By leveraging relative scale estimation from human height and adaptive speed-aware topology, the method outperforms time-based approaches, especially under diverse pedestrian speeds, achieving superior re-identification performance on the SLP dataset with a 20-second average search range.

ABSTRACT

In this paper, we propose a novel distance-based camera network topology inference method for efficient person re-identification. To this end, we first calibrate each camera and estimate relative scales between cameras. Using the calibration results of multiple cameras, we calculate the speed of each person and infer the distance between cameras to generate distance-based camera network topology. The proposed distance-based topology can be applied adaptively to each person according to its speed and handle diverse transition time of people between non-overlapping cameras. To validate the proposed method, we tested the proposed method using an open person re-identification dataset and compared to state-of-the-art methods. The experimental results show that the proposed method is effective for person re-identification in the large-scale camera network with various people transition time.

Motivation & Objective

• To address the challenge of inaccurate person re-identification in large-scale, non-overlapping camera networks due to variable pedestrian transition times.
• To overcome limitations of time-based camera network topologies that assume uniform walking speeds and fail under speed variation.
• To develop a robust, adaptive topology inference method that incorporates spatial distance and individual speed for improved search space restriction.
• To validate the effectiveness of distance-based topology in reducing false positives and improving re-identification accuracy compared to state-of-the-art time-based methods.

Proposed method

• Calibrate cameras using self-calibration techniques to estimate extrinsic parameters up to scale, addressing scale ambiguity in 3D reconstruction.
• Estimate relative scale ratios between cameras using human height information to correct scale ambiguity and improve distance estimation.
• Compute individual person speeds from calibrated camera data and temporal tracking information to infer transition times.
• Calculate inter-camera distances by multiplying person speed with transition time, forming a distance-based camera network topology.
• Dynamically adapt the topology per person by adjusting time-based search ranges based on individual walking speed, enabling speed-aware re-identification.
• Use a consistent baseline for feature extraction and pooling across methods to ensure fair comparison between time-based and distance-based topologies.

Experimental results

Research questions

• RQ1 Can a distance-based camera network topology improve person re-identification performance compared to time-based topology in large-scale camera networks?
• RQ2 How does the proposed method handle diverse pedestrian walking speeds that challenge time-based topology inference?
• RQ3 To what extent does scale estimation from human height improve the accuracy of inter-camera distance inference?
• RQ4 Does the adaptive application of distance-based topology per individual outperform fixed time-based search ranges?

Key findings

• The proposed distance-based method achieved higher retrieval rates than the time-based method, especially in the 5–10 second search range, indicating better performance under variable transition times.
• The method with estimated camera parameters (ours) performed comparably to the ground-truth parameter baseline (gt), demonstrating robustness to calibration errors.
• Even with 25% scale error, the distance-based method outperformed the time-based method, showing resilience to calibration inaccuracies.
• The distance-based topology produced clearer and narrower transition time distributions with smaller variance than time-based methods, indicating more accurate topology inference.
• Despite using a simpler feature pooling method, the proposed distance-based re-identification method outperformed state-of-the-art methods using more complex models like random forests.
• The re-identification performance using the proposed distance-based topology exceeded that of the time-based method, even when both used the same 20-second average search range.