Multi-Camera Industrial Open-Set Person Re-Identification and Tracking

TL;DR

The paper tackles open-set, multi-camera person re-identification in industrial environments under real-time constraints. It introduces MICRO-TRACK, a modular pipeline that fuses single-camera tracking with a selective Re-ID module controlled by a quality threshold, and an Open-Set Gallery that maintains centroid embeddings for identities across time with a circular buffer. The approach achieves robust open-set performance on a real industrial dataset, Facility-ReID, and competitive closed-set results on Market1501 using a ResNet-50 backbone trained with Centroid Triplet Loss, while delivering real-time inference on standard hardware and edge deployments. The Facility-ReID dataset and edge-aware deployment considerations aim to bridge research and practical industrial surveillance needs, with future work targeting overlapping-camera coordination and split-edge computing for scalability.

Abstract

In recent years, the development of deep learning approaches for the task of person re-identification led to impressive results. However, this comes with a limitation for industrial and practical real-world applications. Firstly, most of the existing works operate on closed-world scenarios, in which the people to re-identify (probes) are compared to a closed-set (gallery). Real-world scenarios often are open-set problems in which the gallery is not known a priori, but the number of open-set approaches in the literature is significantly lower. Secondly, challenges such as multi-camera setups, occlusions, real-time requirements, etc., further constrain the applicability of off-the-shelf methods. This work presents MICRO-TRACK, a Modular Industrial multi-Camera Re_identification and Open-set Tracking system that is real-time, scalable, and easy to integrate into existing industrial surveillance scenarios. Furthermore, we release a novel Re-ID and tracking dataset acquired in an industrial manufacturing facility, dubbed Facility-ReID, consisting of 18-minute videos captured by 8 surveillance cameras.

Figures (4)

• Figure 1: A schema of the multi-camera implementation of MICRO-TRACK. A single-camera tracking module extracts the bounding boxes and tracking ids for each camera of the environment. The tracking system considers the occlusions and variability in appearance through time, granting temporal consistency to the re-identification. The decision module (red dot) decides whether the person is qualitatively good enough to extract the Re-ID embeddings. A global Orchestrator system collects all the tracking IDs and Re-ID embeddings, deciding if the person has already been seen by the system or is a new entry. The Orchestrator maintains the open-set gallery.
• Figure 2: Given a sequence of images, a tracking module extracts the tracks for a person, also providing a confidence score on the detection. If the score is greater than a threshold $th_{score}$, the re-identification module is applied, verifying the person's presence in the Open-Set Gallery. If the person is not present, it is considered a new person. Then, the tracking system maintains the person's identity given by the Re-ID module.
• Figure 3: The industrial manufacturing facility's digital twin with the disposition of the multi-camera surveillance system.
• Figure 4: Examples of the images of the Facility-ReID dataset. Each line provides examples for a different camera.