Intel/object-counting

Object Counting

Property	Value
Category	Object Detection + Counting (80-class COCO)
Base Model	YOLO26 (Ultralytics)
Source Framework	PyTorch (Ultralytics)
Supported Precisions	FP32, FP16, INT8 (mixed-precision)
Inference Engine	OpenVINO
Hardware	CPU, GPU, NPU
Detected Class(es)	All 80 COCO classes (counted per class)

---

Overview

Object Counting is a Metro Analytics use case that detects objects and reports how many of each class are present in an image or per video frame. It is built on YOLO26, a state-of-the-art real-time object detector, quantized to INT8 for efficient inference on Intel hardware. Counting is implemented as a thin aggregation layer on top of the strongest general-purpose detector, which keeps it accurate and reusable across classes. The DLStreamer sample below demonstrates this on a traffic scene sample video, counting person, bicycle, and car detections per frame.

Typical Metro deployments include:

• Occupancy Counting -- count people on a platform or in a waiting area.
• Vehicle Counting -- count cars, buses, and trucks at an intersection.
• Inventory Counting -- count bags, bottles, or other items in a zone.
• Throughput Metrics -- aggregate per-frame counts into time series.

Available variants: yolo26n, yolo26s, yolo26m, yolo26l, yolo26x. Smaller variants (yolo26n, yolo26s) are recommended for high-FPS edge deployment; larger variants improve recall for small or distant objects.

For line-crossing counts (directional entry/exit), see the vehicle-entry-exit-logging use case.

---

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 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 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

Replace yolo26n with any variant (yolo26s, yolo26m, yolo26l, yolo26x). The second argument selects the precision (FP32, FP16, INT8); the default is FP16.

The script performs the following steps:

1. Installs dependencies (openvino, ultralytics; adds nncf for INT8).
2. Downloads a sample test image (test.jpg) and a sample test video (test_video.mp4, the person-bicycle-car-detection.mp4 street scene from Intel IoT DevKit's sample-videos repository).
3. Downloads the PyTorch weights and exports to OpenVINO IR.
4. (INT8 only) Quantizes the model using NNCF post-training quantization.

Output files:

• yolo26n_openvino_model/ -- FP32 or FP16 OpenVINO IR model directory.
• yolo26n_objcount_int8.xml / yolo26n_objcount_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, then aggregates detections into a per-class count and a total count for a single image. YOLO26 is end-to-end (NMS-free), so no manual non-maximum suppression is needed. Change the device string to run on CPU, GPU, or NPU.

from collections import Counter

import cv2
import numpy as np
import openvino as ov

CONF_THRESHOLD = 0.4
INPUT_SIZE = 640

core = ov.Core()
model = core.read_model("yolo26n_openvino_model/yolo26n.xml")

# YOLO26 embeds the 80 COCO class names in rt_info -- read them instead of
# hardcoding the list. Ultralytics separates multi-word names with
# underscores (e.g. "traffic_light"), so restore spaces for display.
COCO_NAMES = [
    name.replace("_", " ")
    for name in model.get_rt_info()["model_info"]["labels"].value.split()
]

# Change device to "GPU" or "NPU" to run on integrated GPU or NPU.
compiled = core.compile_model(model, "CPU")

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

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

# YOLO26 end-to-end output: [1, 300, 6] = [x1, y1, x2, y2, confidence, class_id]
output = compiled([blob])[compiled.output(0)][0]
dets = output[output[:, 4] >= CONF_THRESHOLD]

sx, sy = w0 / INPUT_SIZE, h0 / INPUT_SIZE
counts = Counter(COCO_NAMES[int(d[5])] for d in dets)

print(f"Total objects: {len(dets)}")
print("Object counts:")
for name, n in sorted(counts.items(), key=lambda kv: (-kv[1], kv[0])):
    print(f"  {name}: {n}")

colors = np.random.RandomState(42).randint(0, 255, (80, 3)).tolist()
for det in dets:
    x1, y1, x2, y2 = (int(det[0] * sx), int(det[1] * sy),
                      int(det[2] * sx), int(det[3] * sy))
    cid = int(det[5])
    cv2.rectangle(image, (x1, y1), (x2, y2), colors[cid], 2)
    cv2.putText(image, COCO_NAMES[cid], (x1, y1 - 5),
                cv2.FONT_HERSHEY_SIMPLEX, 0.6, colors[cid], 2)

summary = ", ".join(f"{n} {name}" for name, n in counts.items())
cv2.putText(image, summary[:60], (10, 30),
            cv2.FONT_HERSHEY_SIMPLEX, 0.6, (0, 255, 0), 2)
cv2.imwrite("output_openvino.jpg", image)

Device targets:

• "CPU" -- default, works on all Intel platforms.
• "GPU" -- Intel integrated or discrete GPU.
• "NPU" -- Intel NPU (validate with benchmark_app -d NPU).

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 per-class counts to the console, and writes the annotated frame to output_openvino.jpg.

Expected console output (representative):

Total objects: 5
Object counts:
  person: 4
  bus: 1

Expected Output

DLStreamer Sample

The pipeline below runs the FP16 YOLO26 detector on the sample video via gvadetect, overlays bounding boxes with gvawatermark, saves the annotated result to output_dlstreamer.mp4, and prints the per-frame person, bicycle, and car counts by reading the GstAnalytics detection metadata. The sample video (person-bicycle-car-detection.mp4) is a street scene containing pedestrians, a cyclist, and cars, matching the three classes counted below.

Notes on running this sample:

• Use the FP16 IR (yolo26n_openvino_model/yolo26n.xml). Class names are read automatically from the model's embedded metadata.yaml by DLStreamer 2026.0+ -- no external labels-file is required.

• 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:-}
  

from collections import Counter

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"

# Only these classes are counted; the sample video contains pedestrians,
# a cyclist, and cars.
CLASSES_OF_INTEREST = {"person", "bicycle", "car"}

# 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 ! "
    "gvadetect model=yolo26n_openvino_model/yolo26n.xml "
    "device=GPU "
    "threshold=0.4 ! queue ! "
    "gvawatermark ! videoconvert ! video/x-raw,format=I420 ! "
    "openh264enc ! h264parse ! "
    "mp4mux ! filesink name=sink location=output_dlstreamer.mp4"
)
pipeline = Gst.parse_launch(pipeline_str)


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
    counts = Counter()
    idx = 1
    while True:
        ok, od = rmeta.get_od_mtd(idx)
        if not ok:
            break
        label = GLib.quark_to_string(od.get_obj_type())
        if label in CLASSES_OF_INTEREST:
            counts[label] += 1
        idx += 1
    if counts:
        summary = ", ".join(f"{n} {name}" for name, n in counts.items())
        print(f"Object counts: {summary}", flush=True)
    return Gst.PadProbeReturn.OK


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

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.

Expected Output

---

License

Licensed under the MIT License. See LICENSE for details.

References

• YOLO26 Documentation
• Ultralytics Object Counting Guide
• Intel DLStreamer gvadetect
• OpenVINO Documentation
• NNCF Post-Training Quantization
• Intel DLStreamer