Person Re-Identification

Property	Value
Category	Person Detection + Cross-Camera Re-Identification
Base Model	person-detection-retail-0013 + person-reidentification-retail-0287 (Open Model Zoo)
Source Framework	Caffe / PyTorch (Open Model Zoo)
Supported Precisions	FP32, FP16
Inference Engine	OpenVINO
Hardware	CPU, GPU, NPU
Detected Class(es)	Persons (detection) + 256-d appearance embeddings (re-identification)

Overview

Person Re-Identification is a Metro Analytics use case that tracks the same individual across multiple camera views. Given a reference person seen on one camera, it locates that same person on another camera even though the pose, scale, and viewing angle differ. Each detected person is compared to the reference by cosine similarity of its appearance embedding vector.

It uses a two-stage pipeline:

person-detection-retail-0013 -- detects every person in the scene.
person-reidentification-retail-0287 -- computes a 256-d appearance embedding per person that is robust to viewpoint and lighting changes.

Unlike face-based matching, re-identification relies on whole-body appearance (clothing, build, gait cues), so it works at surveillance distances where faces are not clearly visible.

To demonstrate cross-camera behaviour from a single downloadable clip, the wide surveillance video is treated as two virtual cameras by time window: an earlier enrollment window is Camera A (where the reference identity is first seen) and a later query window is Camera B (where the person is re-identified as they continue to move through the scene). This emulates a person first seen on one camera and later re-identified on another using the same embedding-matching logic that links identities across a real multi-camera network.

Typical Metro deployments include:

Multi-Camera Tracking -- follow a person across cameras in campuses, airports, and transit hubs.
Lost-and-Found / Person of Interest -- locate where a flagged individual appears across a camera network.
Journey Analytics -- reconstruct a person's path through a facility.
Tailgating and Zone Analytics -- confirm the same person across entry and interior cameras.

Privacy Note: Person re-identification processes biometric-adjacent appearance data. Ensure your deployment complies with applicable privacy regulations (GDPR, BIPA, etc.) and has proper consent and retention policies in place.

Prerequisites

Python 3.11+
Install OpenVINO (latest version)
Install Intel DLStreamer (latest version)

Create and activate a Python virtual environment before running the scripts:

python3 -m venv .venv --system-site-packages
source .venv/bin/activate

Note: The --system-site-packages flag is required so the virtual environment can access the system-installed OpenVINO and DLStreamer Python packages.

Getting Started

Download Models

Run the provided script to download the person detection and re-identification models from the Open Model Zoo:

chmod +x export_and_quantize.sh
./export_and_quantize.sh

The script downloads person-detection-retail-0013 and person-reidentification-retail-0287 in FP16, downloads the sample surveillance video (test_video.mp4), and captures a reference person crop (person_a.jpg) of the most prominent person seen in the Camera A enrollment window.

OpenVINO Sample

The sample below re-identifies the reference person across camera views. It loads the captured reference image (person_a.jpg, enrolled from Camera A), computes its embedding, then scans frames of the Camera B query window. In each Camera B frame it detects every person, embeds each one, and keeps the person whose similarity to the reference is highest. It writes the Camera B frame with the strongest match, drawing a green box only on the re-identified person. Change the device string to run on CPU, GPU, or NPU.

import cv2
import numpy as np
import openvino as ov

DETECTION_MODEL = "intel/person-detection-retail-0013/FP16/person-detection-retail-0013.xml"
REID_MODEL = "intel/person-reidentification-retail-0287/FP16/person-reidentification-retail-0287.xml"
REFERENCE_IMAGE = "person_a.jpg"     # reference identity enrolled from Camera A
SCENE_VIDEO = "test_video.mp4"       # wide feed; a later window acts as Camera B
CAMERA_B_START_FRAME = 450           # query window begins ~15s into the clip
CONF_THRESHOLD = 0.6
MATCH_THRESHOLD = 0.6

core = ov.Core()

# Change device to "GPU" or "NPU" to run on integrated GPU or NPU.
det_compiled = core.compile_model(core.read_model(DETECTION_MODEL), "CPU")
reid_compiled = core.compile_model(core.read_model(REID_MODEL), "CPU")

det_input = det_compiled.input(0)
det_h, det_w = det_input.shape[2], det_input.shape[3]
reid_input = reid_compiled.input(0)
reid_h, reid_w = reid_input.shape[2], reid_input.shape[3]


def detect_persons(img):
    h0, w0 = img.shape[:2]
    blob = cv2.resize(img, (det_w, det_h))
    blob = blob.transpose(2, 0, 1)[np.newaxis, ...].astype(np.float32)
    dets = det_compiled([blob])[det_compiled.output(0)][0][0]
    persons = []
    for d in dets:
        if float(d[2]) < CONF_THRESHOLD:
            continue
        x1 = max(0, int(d[3] * w0))
        y1 = max(0, int(d[4] * h0))
        x2 = min(w0, int(d[5] * w0))
        y2 = min(h0, int(d[6] * h0))
        if x2 > x1 and y2 > y1:
            persons.append((x1, y1, x2, y2))
    return persons


def get_embedding(img, bbox):
    x1, y1, x2, y2 = bbox
    crop = img[y1:y2, x1:x2]
    blob = cv2.resize(crop, (reid_w, reid_h))
    blob = blob.transpose(2, 0, 1)[np.newaxis, ...].astype(np.float32)
    emb = reid_compiled([blob])[reid_compiled.output(0)].flatten()
    return emb / (np.linalg.norm(emb) + 1e-9)


# 1. Embed the reference person enrolled from Camera A.
# person_a.jpg is already a cropped person, so embed the whole image directly
# (the re-identification model expects a person crop as its input).
reference = cv2.imread(REFERENCE_IMAGE)
if reference is None:
    raise SystemExit("Could not read the reference image")
ref_emb = get_embedding(reference, (0, 0, reference.shape[1], reference.shape[0]))

# 2. Scan the Camera B window and keep the frame with the strongest re-id match.
cap = cv2.VideoCapture(SCENE_VIDEO)
total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT)) or 900
best = {"sim": 0.0, "frame": None, "bbox": None}
for frame_idx in range(CAMERA_B_START_FRAME, total, 15):
    cap.set(cv2.CAP_PROP_POS_FRAMES, frame_idx)
    ok, frame = cap.read()
    if not ok:
        break
    for bbox in detect_persons(frame):
        sim = float(np.dot(get_embedding(frame, bbox), ref_emb))
        if sim > best["sim"]:
            best = {"sim": sim, "frame": frame.copy(), "bbox": bbox}
cap.release()

# 3. Annotate and save the best Camera B match.
if best["frame"] is None:
    raise SystemExit("No person detected in the Camera B window")
frame = best["frame"]
if best["sim"] >= MATCH_THRESHOLD:
    x1, y1, x2, y2 = best["bbox"]
    cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0), 2)
    cv2.putText(frame, f"RE-ID {best['sim']:.2f}", (x1, max(15, y1 - 8)),
                cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)
    print(f"Re-identified reference person in Camera B, similarity {best['sim']:.4f}")
else:
    print(f"No matching person found (best similarity {best['sim']:.4f})")

cv2.imwrite("output_openvino.jpg", frame)
print("Saved: output_openvino.jpg")

Device targets:

"CPU" -- default, works on all Intel platforms.
"GPU" -- Intel integrated or discrete GPU.
"NPU" -- Intel NPU; both FP16 models are NPU-compatible.

Expected Output

DLStreamer Sample

The pipeline below runs the person detector via gvadetect and the re-identification model via gvaclassify on the video. Frames are pulled through an appsink, where each detected person's embedding is compared to the reference embedding computed from person_a.jpg. Only persons that match the reference identity are boxed, so the annotated output_dlstreamer.mp4 highlights the same person as they move across the scene even when other people are present.

Notes on running this sample:
Export PYTHONPATH so the DLStreamer Python modules (gi, gstgva) are importable:
source /opt/intel/openvino_2026/setupvars.sh
source /opt/intel/dlstreamer/scripts/setup_dls_env.sh
export PYTHONPATH=/opt/intel/dlstreamer/python:\
/opt/intel/dlstreamer/gstreamer/lib/python3/dist-packages:${PYTHONPATH:-}
The re-identification embedding is attached as a tensor on each person's region-of-interest metadata. Convert the stream to BGR before gvadetect/gvaclassify so a downstream format conversion does not strip those tensors before the appsink reads them.

import gi

gi.require_version("Gst", "1.0")
from gi.repository import Gst

Gst.init([])

import numpy as np
import cv2
from gstgva import VideoFrame

INPUT_VIDEO = "test_video.mp4"
REFERENCE_IMAGE = "person_a.jpg"
OUTPUT_VIDEO = "output_dlstreamer.mp4"
DETECTION_MODEL = "intel/person-detection-retail-0013/FP16/person-detection-retail-0013.xml"
REID_MODEL = "intel/person-reidentification-retail-0287/FP16/person-reidentification-retail-0287.xml"
# For CPU: change "GPU" to "CPU". For NPU: change "GPU" to "NPU".
DEVICE = "GPU"
DET_THRESHOLD = 0.6
MATCH_THRESHOLD = 0.6


def person_embeddings(video_frame):
    """Yield ((x, y, w, h), normalized_embedding) for each classified person."""
    for region in video_frame.regions():
        rect = region.rect()
        emb = None
        for tensor in region.tensors():
            if tensor.is_detection():
                continue
            data = np.array(tensor.data(), dtype=np.float32)
            if data.size >= 256:
                emb = data[:256]
        if emb is None:
            continue
        emb = emb / (np.linalg.norm(emb) + 1e-9)
        yield (int(rect.x), int(rect.y), int(rect.w), int(rect.h)), emb


def run_pipeline(source_desc, on_frame):
    # Convert to BGR before inference so gvaclassify's embedding tensors survive
    # to the appsink (a later format-changing videoconvert would strip them).
    pipeline = Gst.parse_launch(
        f"{source_desc} ! videoconvert ! video/x-raw,format=BGR ! "
        f"gvadetect model={DETECTION_MODEL} device={DEVICE} "
        f"threshold={DET_THRESHOLD} ! queue ! "
        f"gvaclassify model={REID_MODEL} device={DEVICE} ! queue ! "
        "appsink name=sink emit-signals=true sync=false max-buffers=4 drop=false"
    )
    sink = pipeline.get_by_name("sink")
    sink.connect("new-sample", on_frame)
    pipeline.set_state(Gst.State.PLAYING)
    pipeline.get_bus().timed_pop_filtered(
        Gst.CLOCK_TIME_NONE, Gst.MessageType.EOS | Gst.MessageType.ERROR)
    pipeline.set_state(Gst.State.NULL)


# 1. Compute the reference embedding from the enrolled person image.
# person_a.jpg is already a person crop, so run the re-identification model on
# the whole frame (inference-region=full-frame) instead of detecting first.
ref = {"emb": None}


def on_reference(sink):
    sample = sink.emit("pull-sample")
    if sample is None:
        return Gst.FlowReturn.OK
    vf = VideoFrame(sample.get_buffer(), caps=sample.get_caps())
    for tensor in vf.tensors():
        data = np.array(tensor.data(), dtype=np.float32)
        if data.size >= 256:
            emb = data[:256]
            ref["emb"] = emb / (np.linalg.norm(emb) + 1e-9)
    return Gst.FlowReturn.OK


reference_pipeline = Gst.parse_launch(
    f"filesrc location={REFERENCE_IMAGE} ! jpegdec ! videoconvert ! "
    f"video/x-raw,format=BGR ! "
    f"gvainference model={REID_MODEL} device={DEVICE} "
    f"inference-region=full-frame ! queue ! "
    "appsink name=sink emit-signals=true sync=false max-buffers=4 drop=false"
)
ref_sink = reference_pipeline.get_by_name("sink")
ref_sink.connect("new-sample", on_reference)
reference_pipeline.set_state(Gst.State.PLAYING)
reference_pipeline.get_bus().timed_pop_filtered(
    Gst.CLOCK_TIME_NONE, Gst.MessageType.EOS | Gst.MessageType.ERROR)
reference_pipeline.set_state(Gst.State.NULL)
if ref["emb"] is None:
    raise SystemExit("Could not compute the reference embedding")
ref_emb = ref["emb"]

# 2. Process the video, boxing only persons that match the reference identity.
writer = {"w": None}
match_frames = 0


def on_video(sink):
    global match_frames
    sample = sink.emit("pull-sample")
    if sample is None:
        return Gst.FlowReturn.OK
    vf = VideoFrame(sample.get_buffer(), caps=sample.get_caps())
    matches = []
    for (x, y, w, h), emb in person_embeddings(vf):
        similarity = float(np.dot(emb, ref_emb))
        if similarity >= MATCH_THRESHOLD:
            matches.append((x, y, w, h, similarity))

    with vf.data() as mat:
        frame = mat.copy()

    if writer["w"] is None:
        frame_h, frame_w = frame.shape[:2]
        structure = sample.get_caps().get_structure(0)
        ok_fr, fps_n, fps_d = structure.get_fraction("framerate")
        fps = fps_n / fps_d if ok_fr and fps_d else 12
        writer["w"] = cv2.VideoWriter(
            OUTPUT_VIDEO, cv2.VideoWriter_fourcc(*"mp4v"), fps, (frame_w, frame_h))

    for x, y, w, h, similarity in matches:
        cv2.rectangle(frame, (x, y), (x + w, y + h), (0, 255, 0), 2)
        cv2.putText(frame, f"RE-ID {similarity:.2f}", (x, max(15, y - 8)),
                    cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)
    if matches:
        match_frames += 1
    writer["w"].write(frame)
    return Gst.FlowReturn.OK


run_pipeline(f"filesrc location={INPUT_VIDEO} ! decodebin3", on_video)
if writer["w"] is not None:
    writer["w"].release()
print(f"Frames with a re-identified person: {match_frames}", flush=True)
print(f"Saved: {OUTPUT_VIDEO}", flush=True)

Device targets:

DEVICE = "GPU" -- default in the sample code.
DEVICE = "CPU" -- change "GPU" to "CPU".
DEVICE = "NPU" -- change "GPU" to "NPU"; use batch-size=1 and nireq=4 for best NPU utilization.

Expected Output

License

Licensed under the MIT License. See LICENSE for details.

References

Downloads last month: -; Downloads are not tracked for this model. How to track

Collection including Intel/person-reidentification

Metro Analytics Catalog

Collection

AI models and analytics pipelines for Intel hardware to help developers build and deploy edge AI solutions quickly. • 27 items • Updated 4 days ago • 1