Consent, Privacy & Filming People

Introduction

Your next field trip films real people playing football. A photo of a person isn't just a picture: it's personal data, and the law and basic respect both say you must ask first, collect only what you need, and look after it. Today you'll learn what that means in practice, write your team's privacy rules, and build a script that blurs faces.

By the end of this lesson you'll be able to:

• Explain why an image of a person is personal data, not just "a picture"
• Describe what informed consent means, and why it differs for children
• Say, in plain words, what India's DPDP Act 2023 and the GDPR ask of anyone who collects images
• Decide when a venue permission letter is needed, and what it must say
• Apply data minimisation: collect only what the project needs, keep it only as long as needed
• Use a script to blur faces in stored images
• Write a one-page team privacy and consent charter

Working principle

Important: this lesson is a teaching summary, not legal advice. Laws change and depend on your school, state and country. Ask your school administration or a lawyer to check your letters before you use them.

A photo of a person is personal data

If a person can be recognised from an image (by their face, their jersey number, their voice in a video, or by the context), the image is personal data. It is not just pixels. That means someone is responsible for it: you.

The four questions before filming anyone

What "consent" has to look like

Good consent is:

• Informed: people are told what you collect, why, who sees it and how long you keep it
• Specific: one purpose, not "anything we like"
• Freely given: saying no must be easy and without penalty
• Clear: a signature, a ticked box or a spoken yes on record, not silence
• Withdrawable: a person can change their mind, and you delete their images

For children, the consent comes from a parent or guardian, and the child should still be told in words they understand and may say no.

Filming a crowd or a public place

"It was in a public place" doesn't mean "anything goes". Public place or not, think about purpose and harm.

The law, in plain words

India: Digital Personal Data Protection Act, 2023 (DPDP). The Act covers personal data processed digitally. In broad terms it asks the organisation that decides why and how data is used (the "data fiduciary") to process data for a lawful purpose, usually with consent, to give a clear notice, to keep data safe, and to delete it when no longer needed. For children (under 18) it requires verifiable parental consent, and it bars tracking or targeted advertising aimed at children. The Rules were notified on 13 November 2025 and switch on in phases: the first part immediately, consent managers about a year later (13 November 2026), and the substantive obligations, including those for children's data and the penalties, about 18 months after notification (13 May 2027). There are exemptions, including some for educational institutions when processing is limited to education, health or safety. Whether training an AI model counts as one of those purposes is not settled.

European Union: GDPR. If you ever collect images of people in the EU, the GDPR applies. It asks for a lawful basis (often consent), clear information, data minimisation, a limit on how long you store data, and the right to have data deleted. The age at which a child can consent alone is set by each country, between 13 and 16.

What we do: because the rules are still moving, we follow the stricter, simpler standard now: ask first, collect little, protect it, delete it, and never use it for something people didn't agree to. Check the latest position before you launch anything real.

Data minimisation: the best privacy tool

What you'll need

• Your cv_course environment, for the face-blurring script
• A few test photos (of yourself, or ones you have permission to use)

Build it

Step 1: blur faces in a folder of images

OpenCV can find faces in two ways, depending on the version you installed (check with python -c "import cv2; print(cv2.__version__)"):

• OpenCV 4.x: the built-in Haar face detector (cv2.CascadeClassifier). Nothing to download.
• OpenCV 5.x: that classic detector was removed from the main package. Use the YuNet face model instead. Download face_detection_yunet_2023mar.onnx from the OpenCV Zoo (github.com/opencv/opencv_zoo, folder models/face_detection_yunet) and put it next to your script.

The script below tries both. Neither is perfect: faces that are small, turned away or in shadow get missed. So this is a safety net, not a guarantee. Always check the output by eye.

import os
 
import cv2
 
YUNET_MODEL = "face_detection_yunet_2023mar.onnx"      # needed for OpenCV 5.x only
 
 
def make_face_finder():
    """Return a function find_faces(image) -> list of (x, y, w, h)."""
    if hasattr(cv2, "CascadeClassifier"):                   # OpenCV 4.x
        cascade = cv2.CascadeClassifier(cv2.data.haarcascades + "haarcascade_frontalface_default.xml")
 
        def find_faces(image):
            gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
            found = cascade.detectMultiScale(gray, scaleFactor=1.1, minNeighbors=5, minSize=(24, 24))
            return [tuple(int(v) for v in f) for f in found]
 
        return find_faces
 
    if not os.path.exists(YUNET_MODEL):                     # OpenCV 5.x
        raise SystemExit(f"Download {YUNET_MODEL} from the OpenCV Zoo and put it in this folder.")
    yunet = cv2.FaceDetectorYN_create(YUNET_MODEL, "", (320, 320), 0.6, 0.3, 5000)
 
    def find_faces(image):
        h, w = image.shape[:2]
        yunet.setInputSize((w, h))
        _, found = yunet.detect(image)
        if found is None:
            return []
        return [tuple(int(v) for v in row[:4]) for row in found]
 
    return find_faces
 
 
def blur_faces(image, find_faces):
    """Return (blurred copy, number of faces found)."""
    faces = find_faces(image)
    out = image.copy()
    for (x, y, w, h) in faces:
        pad = int(0.15 * w)                                   # blur a bit beyond the face box
        x1, y1 = max(0, x - pad), max(0, y - pad)
        x2, y2 = min(image.shape[1], x + w + pad), min(image.shape[0], y + h + pad)
        if x2 <= x1 or y2 <= y1:
            continue
        k = max(15, ((x2 - x1) // 3) | 1)                     # strong, odd kernel
        out[y1:y2, x1:x2] = cv2.GaussianBlur(out[y1:y2, x1:x2], (k, k), 0)
    return out, len(faces)
 
 
def blur_folder(src, dst):
    find_faces = make_face_finder()
    os.makedirs(dst, exist_ok=True)
    total_faces = 0
    for name in sorted(os.listdir(src)):
        if not name.lower().endswith((".jpg", ".jpeg", ".png")):
            continue
        img = cv2.imread(os.path.join(src, name))
        if img is None:
            continue
        blurred, n = blur_faces(img, find_faces)
        cv2.imwrite(os.path.join(dst, name), blurred)
        total_faces += n
        if n:
            print(f"{name}: blurred {n} face(s)")
    print("Done. Faces blurred:", total_faces, "-> now CHECK the images by eye.")
 
 
if __name__ == "__main__":
    blur_folder("dataset/player", "dataset_blurred/player")

Question: these detectors find faces best when they look towards the camera. Which faces will it miss on a football pitch? What else could you do to protect those people?

Step 2: keep a consent log (no personal details in the file)

A tiny CSV that records that consent exists, without copying anyone's data around.

import csv
import os
from datetime import date
 
LOG = "consent_log.csv"
FIELDS = ["form_id", "group", "consent_given", "date_received", "delete_by", "notes"]
 
 
def add_entry(form_id, group, consent_given, delete_by, notes=""):
    new = not os.path.exists(LOG)
    with open(LOG, "a", newline="") as f:
        w = csv.DictWriter(f, fieldnames=FIELDS)
        if new:
            w.writeheader()
        w.writerow({"form_id": form_id, "group": group, "consent_given": consent_given,
                    "date_received": date.today().isoformat(), "delete_by": delete_by, "notes": notes})
 
 
def summary():
    with open(LOG) as f:
        rows = list(csv.DictReader(f))
    yes = [r for r in rows if r["consent_given"] == "yes"]
    no = [r for r in rows if r["consent_given"] != "yes"]
    print(f"{len(yes)} consent(s) on file, {len(no)} refused or missing")
    return rows
 
 
add_entry("F001", "U14 team A", "yes", "2027-03-31")
add_entry("F002", "U14 team A", "no", "2027-03-31", "do not film this child, keep out of frame")
summary()

Question: why does the log use a form number and not the child's name?

Case study cards

Each is fictional and written for discussion.

Templates

Consent form (parents or guardians): sample wording

Project: AI and Computer Vision course, football data collection What we will do: Students will film a football session with a laptop webcam to build a small image dataset, and label the players and ball in the images. What we collect: Images of players (including children under 18) from the pitch side. We will not record sound. We will not record names. Why: To teach students how to build and test an object-detection model. The images will not be used for advertising, tracking, selling, or any other purpose. Who sees it: The students on the team, their teacher, and, for labelling, the online tool __________ (stores data in __________). How long we keep it: Until __ / __ / ____, then we delete all raw images and labelled copies. Faces: We will blur faces before sharing images outside the team: Yes / No. Your rights: You may refuse without any consequence for your child. You may withdraw consent at any time and ask us to delete your child's images. Contact __________. Consent: I am the parent or guardian of __________. I agree to my child being filmed for this purpose. Signature __________ Date __________

Venue permission letter: sample wording

To: The Manager or Owner, __________ (ground or club) From: __________ (school), course teacher __________

We request permission for __ students and __ teacher(s) to set up a laptop and webcam at __________ on __ / __ / ____ between __:__ and __:__ to film football play for an educational data-collection exercise.

We will keep to the venue's rules, avoid blocking any area, point the camera at the playing area and not at spectators, stop filming anyone who asks us to, and delete the raw footage after the project. We can share our consent form and privacy charter on request.

Please sign below to confirm permission. Name, position, signature, date.

Team Privacy and Consent Charter (one page)

If it goes wrong

Try it

• Rewrite the parent consent form in language a 10-year-old can understand
• Research one more privacy law (for example the Californian CCPA or the UK GDPR) and compare it in 5 lines
• Extend blur_folder to also blur jersey numbers in a marked region

Conclusion

Images of people are personal data. Before filming, know your purpose, get informed consent (from parents for children), handle bystanders and venues, and follow the law (India's DPDP Act, the GDPR). The best privacy tool is collecting less: film from a distance, blur faces, use ids instead of names, and delete what you don't need.

What next?

Next: Field trip prep: data collection plan. You'll plan the first field trip with everything from Unit 3.