Distance-based Camera Network Topology Inference for Person Re-identification

TL;DR

This work tackles large-scale person re-identification across networks of non-overlapping cameras where diverse pedestrian speeds hinder traditional time-based topology. It introduces a distance-based camera network topology by automatically estimating relative camera scales through height-based self-calibration, extracting per-identity speeds, and constructing a distance distribution $p^{kl}(d)$ between camera pairs. Re-identification is performed within an adaptive search range derived from per-identity transition times, using pooled appearance features and a simple exponential similarity, $S(o^k_i,o^l_j)=\exp(-\|\Phi(o^k_i)-\Phi(o^l_j)\|_2)$. Experiments on the synchronized SLP dataset show that the distance-based topology produces sharper distance distributions and higher retrieval and rank-1 accuracy than time-based approaches, with robustness to calibration errors. Overall, the method enables scalable, accurate re-identification in large camera networks by fusing geometry with appearance cues and adapting search ranges to individual speeds.

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.

Figures (11)

• Figure 1: Challenges in person re-identification based on the time-based camera network topology due to the diverse speeds of people. Each blob was marked at every 0.3 seconds interval and each color indicates a person identity.
• Figure 2: Overview of the proposed framework: distance-based camera network topology inference and person re-identification.
• Figure 3: Examples of detected human heights in a 2D image and corresponding 3D human heights in a world coordinate system.
• Figure 4: Example of estimating a distance between two cameras. The speed in the blind area is inferred by averaging two speeds from two cameras. The distance between cameras is estimated as $46m$ from both identities.
• Figure 5: Comparison of two distributions between cameras. The transition time distribution has a larger variance than that of the distance distribution.