Intel/crowd-detection

Crowd Detection

Property	Value
Category	Object Detection (Crowd / Person Counting)
Base Model	YOLO26 (Ultralytics, default) or DETR (Facebook, alternative)
Source Framework	PyTorch (Ultralytics, Hugging Face Transformers)
Supported Precisions	FP32, FP16, INT8 (mixed-precision)
Inference Engine	OpenVINO
Hardware	CPU, GPU, NPU
Detected Class	person (COCO class 0)

---

Overview

Crowd Detection is a Metro Analytics use case that detects and counts people in video streams to estimate occupancy and identify crowd build-up. It is built on YOLO26, a state-of-the-art real-time object detector trained on the COCO dataset, quantized to INT8 and filtered at runtime to the person class. Typical Metro deployments include:

• Platform Occupancy -- count waiting passengers on station platforms.
• Entry / Exit Flow -- monitor pedestrian throughput at gates and turnstiles.
• Crowd Build-up Alerts -- trigger notifications when person counts cross a threshold.
• Public Safety Analytics -- support situational awareness in transit hubs and venues.

Available variants: yolo26n, yolo26s, yolo26m, yolo26l. Smaller variants (yolo26n, yolo26s) are recommended for high-FPS edge deployment; larger variants improve recall in dense crowds.

Model Selection

DETR (DEtection TRansformer) is available as an alternative to YOLO26. DETR is a transformer-based, NMS-free detector trained on COCO and released under the Apache-2.0 license. The export script embeds DETR normalization, box decoding, and the class remap into the OpenVINO IR, so the DETR model produces the same output layout and class ids as YOLO26. The same OpenVINO and DLStreamer samples run either model; select one with the MODEL constant at the top of each sample.

Model	Variants	Input Size	Output	License
YOLO26 (default)	yolo26n, yolo26s, yolo26m, yolo26l	640 x 640	[1, 300, 6]	AGPL-3.0
DETR	detr-resnet-50	800 x 800	[1, 100, 6]	Apache-2.0

Choose YOLO26 for maximum throughput on edge devices. Choose DETR when a permissively licensed transformer detector is preferred; it is heavier and runs at a lower frame rate.

---

Prerequisites

Target stack: Intel DLStreamer 2026.3.0 RC1 with the matching OpenVINO GenAI bundle 2026.5.0.0.dev20261030 (OpenVINO Runtime 2026.5.0.dev20261030). Use the native RC1 setup before running these samples; this is a development stack, not a stable release. See the device validation matrix for measured CPU, GPU, and NPU results.

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

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 and Quantize Model

Run the provided script to download, export to OpenVINO IR, and optionally quantize:

chmod +x export_and_quantize.sh
./export_and_quantize.sh

This exports the default yolo26n model in FP16 precision.

Optional: Select a Different Model, Variant, or Precision

./export_and_quantize.sh yolo26n FP32          # full-precision
./export_and_quantize.sh yolo26n INT8          # quantized
./export_and_quantize.sh yolo26s               # larger variant, default FP16
./export_and_quantize.sh detr-resnet-50        # DETR alternative, default FP16
./export_and_quantize.sh detr-resnet-50 INT8   # DETR quantized

The first argument selects the model: any YOLO26 variant (yolo26n, yolo26s, yolo26m, yolo26l) or DETR (detr-resnet-50). The second argument selects the precision (FP32, FP16, INT8); the default is FP16.

The script performs the following steps:

1. Installs dependencies (openvino, ultralytics for YOLO26 or transformers, torch for DETR; adds nncf for INT8).
2. Downloads a sample test image (test.jpg) and a sample test video (test_video.mp4).
3. Downloads the PyTorch weights and exports to OpenVINO IR.
4. (INT8 only) Quantizes the model using NNCF post-training quantization.

Output files (replace yolo26n with the selected model, for example detr-resnet-50):

• yolo26n_openvino_model/ -- FP32 or FP16 OpenVINO IR model directory.
• yolo26n_crowd_int8.xml / yolo26n_crowd_int8.bin -- INT8 quantized model (only when INT8 is selected).

Precision / Device Compatibility

Precision	CPU	GPU	NPU
FP32	Yes	Yes	No
FP16	Yes	Yes	Yes
INT8	Yes	Yes	Yes

Note: The INT8 calibration uses the bundled sample image. For production accuracy, replace it with a representative set of frames from the target deployment site.

OpenVINO Sample

The sample below runs YOLO26 inference, filters to the person class, and reports the crowd count for a single image.

The full sample is in openvino_sample.py; run it with python3 openvino_sample.py or copy the code below.

# SPDX-License-Identifier: MIT
# Copyright (C) Intel Corporation
#
# Crowd Detection: OpenVINO Runtime sample.
# See README.md for prerequisites and inputs. Run: python3 openvino_sample.py

import cv2
import numpy as np
import openvino as ov

PERSON_CLASS_ID = 0
CONF_THRESHOLD = 0.4
# Select the model: "yolo26n" (default) or "detr-resnet-50".
MODEL = "yolo26n"
INPUT_SIZE = 800 if MODEL.startswith("detr") else 640

core = ov.Core()
model = core.read_model(f"{MODEL}_openvino_model/{MODEL}.xml")
compiled = core.compile_model(model, "CPU")  # or "GPU", "NPU"

image = cv2.imread("test.jpg")
h0, w0 = image.shape[:2]

# Preprocess: letterbox-free resize for simplicity.
blob = cv2.resize(image, (INPUT_SIZE, INPUT_SIZE))
blob = cv2.cvtColor(blob, cv2.COLOR_BGR2RGB).astype(np.float32) / 255.0
blob = blob.transpose(2, 0, 1)[np.newaxis, ...]  # NCHW

# End-to-end output: [1, N, 6] = [x1, y1, x2, y2, confidence, class_id]
# (N = 300 for YOLO26, 100 for DETR)
# No NMS is needed -- YOLO26 is natively end-to-end.
output = compiled([blob])[compiled.output(0)][0]
mask = (output[:, 4] >= CONF_THRESHOLD) & (output[:, 5].astype(int) == PERSON_CLASS_ID)
dets = output[mask]

sx, sy = w0 / INPUT_SIZE, h0 / INPUT_SIZE
crowd_count = len(dets)
print(f"Detected persons: {crowd_count}")

for det in dets:
    x1 = int(det[0] * sx)
    y1 = int(det[1] * sy)
    x2 = int(det[2] * sx)
    y2 = int(det[3] * sy)
    cv2.rectangle(image, (x1, y1), (x2, y2), (0, 255, 0), 2)

cv2.putText(
    image, f"Crowd count: {crowd_count}", (10, 30),
    cv2.FONT_HERSHEY_SIMPLEX, 1.0, (0, 255, 0), 2,
)
cv2.imwrite("output_openvino.jpg", image)

Try It on a Sample Image

The export_and_quantize.sh script downloads test.jpg automatically. Re-run the OpenVINO sample above. The script reads test.jpg, prints the crowd count to the console, and writes the annotated frame to output_openvino.jpg.

Expected console output:

Detected persons: 4

output_openvino.jpg is the same image with a green bounding box drawn around each detected person and the text Crowd count: 4 overlaid in the top-left corner.

Tip: For production testing, replace the bundled test.jpg with an image from your target deployment site showing a representative crowd density.

Expected Output

DLStreamer Sample

The pipeline below runs the FP16 YOLO26 detector on the sample video via gvadetect, filters detections to the person class in a buffer probe using the GStreamer Analytics metadata API (GstAnalytics), overlays bounding boxes, saves the annotated result to output_dlstreamer.mp4, and prints the crowd count per frame.

Notes on running this sample:

• Use the FP16 IR (yolo26n_openvino_model/yolo26n.xml or detr-resnet-50_openvino_model/detr-resnet-50.xml). The DETR IR carries the same model_info metadata as YOLO26, so gvadetect uses the same YOLO26 post-processor for both models. gvadetect selects the YOLO26 post-processor from the metadata.yaml file stored next to the model. To use the INT8 model, place it in a directory together with that file, for example mkdir int8 && cp yolo26n_crowd_int8.* yolo26n_openvino_model/metadata.yaml int8/, and set model=int8/yolo26n_crowd_int8.xml.

• Class names are read automatically from the model's embedded metadata.yaml by DLStreamer 2026.0+ -- no external labels-file is required.

• Filtering with object-class=person directly on gvadetect is rejected when inference-region is full-frame (the default), so the sample filters by detection label in the buffer probe instead.

• Export PYTHONPATH so the DLStreamer Python module is 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 full sample is in dlstreamer_sample.py; run it with python3 dlstreamer_sample.py or copy the code below.

# SPDX-License-Identifier: MIT
# Copyright (C) Intel Corporation
#
# Crowd Detection: DLStreamer pipeline sample.
# See README.md for prerequisites and inputs. Run: python3 dlstreamer_sample.py

import gi

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

Gst.init([])

INPUT_VIDEO = "test_video.mp4"

# Select the model: "yolo26n" (default) or "detr-resnet-50".
MODEL = "yolo26n"

# For CPU: change device=GPU to device=CPU.
# For NPU: change device=GPU to device=NPU (batch-size=1, nireq=4 recommended).
pipeline_str = (
    f"filesrc location={INPUT_VIDEO} ! decodebin3 ! "
    "videoconvert ! "
    f"gvadetect model={MODEL}_openvino_model/{MODEL}.xml "
    "device=GPU "
    "threshold=0.4 ! queue ! "
    "gvawatermark displ-cfg=show-roi=person ! "
    "videoconvert ! video/x-raw,format=I420 ! "
    "openh264enc ! h264parse ! "
    "mp4mux ! filesink name=sink location=output_dlstreamer.mp4"
)
pipeline = Gst.parse_launch(pipeline_str)

sink = pipeline.get_by_name("sink")
sink_pad = sink.get_static_pad("sink")


def on_buffer(pad, info):
    buf = info.get_buffer()
    rmeta = GstAnalytics.buffer_get_analytics_relation_meta(buf)
    if rmeta is None:
        return Gst.PadProbeReturn.OK
    crowd_count = 0
    idx = 1
    while True:
        ok, od = rmeta.get_od_mtd(idx)
        if not ok:
            break
        if GLib.quark_to_string(od.get_obj_type()) == "person":
            crowd_count += 1
        idx += 1
    if crowd_count:
        print(f"Crowd count: {crowd_count}", flush=True)
    return Gst.PadProbeReturn.OK


sink_pad.add_probe(Gst.PadProbeType.BUFFER, on_buffer)

pipeline.set_state(Gst.State.PLAYING)
bus = pipeline.get_bus()
bus.timed_pop_filtered(
    Gst.CLOCK_TIME_NONE,
    Gst.MessageType.EOS | Gst.MessageType.ERROR,
)
pipeline.set_state(Gst.State.NULL)

Expected Output

Device targets:

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

---

License

Licensed under the MIT License. See LICENSE for details.

References

• YOLO26 Documentation
• DETR Model Card (facebook/detr-resnet-50)
• End-to-End Object Detection with Transformers (DETR paper)
• OpenVINO YOLO26 Notebook
• COCO Dataset
• OpenVINO Documentation
• NNCF Post-Training Quantization
• Intel DLStreamer