Automatic Face Ban: A Privacy-Focused Streaming Solution for IoT

This project demonstrates how to automatically censor (ban) faces in a live video stream to ensure privacy using the Portenta H7 and the Portenta Vision Shield.

Jan 28, 2025

•

3127 views

•

0 respects

Components and supplies

Portenta H7

Portenta Vision Shield - LoRa®

Portenta Vision Shield - Ethernet

Apps and platforms

OpenMV IDE

Project description

Code

Face Ban (Ethernet)

python

The Ethernet version of the face ban script.

1"""
2Automatic Face Ban (Ethernet Version)
3Name: portenta_faceban_ethernet.py
4Purpose:
5    This script demonstrates how to automatically censor (ban) faces in
6    a live video stream for privacy protection using an Ethernet connection.
7    Using TensorFlow Lite FOMO (Fast Object Detection Model) on a 
8    Portenta H7 + Portenta Vision Shield, faces are detected in real time 
9    and an overlay is used to obscure them before streaming via MJPEG over HTTP.
10    
11    By enlarging the overlay and keeping it on-screen for a short hold time
12    (even if the face momentarily disappears), this solution helps ensure
13    that no identifiable faces are exposed, which is crucial for compliance
14    with privacy regulations in environments like schools, hospitals,
15    or public buildings.
16
17Author: Arduino Product Experience Team
18Version: 1.0 (27/01/25)
19
20Key Features:
21    - MJPEG video streaming over HTTP for real-time viewing.
22    - Automatic face detection and overlay (face ban/censorship).
23    - Configurable hold time to keep obscuring the face after detection loss.
24    - Ideal for IoT-edge deployments where on-device processing is required.
25"""
26
27import sensor
28import time
29import image
30import network
31import socket
32import ml
33from ml.utils import NMS
34import math
35
36# ------------------------------------------------------------------------------
37# User Configuration
38# ------------------------------------------------------------------------------
39HOST = ""  # Binds to the first available interface
40PORT = 8080
41
42# ------------------------------------------------------------------------------
43# Detection Parameters
44# ------------------------------------------------------------------------------
45MIN_CONFIDENCE = 0.4           # Minimum confidence threshold for faces
46ENLARGEMENT_FACTOR = 3.5       # Scale factor to enlarge the overlay over the face
47HOLD_TIME_MS = 1500            # How long (ms) to keep overlay after losing detection
48
49# ------------------------------------------------------------------------------
50# Camera Setup
51# ------------------------------------------------------------------------------
52sensor.reset()
53sensor.set_pixformat(sensor.GRAYSCALE)
54sensor.set_framesize(sensor.QVGA)
55sensor.set_windowing((240, 240))  # Typical FOMO model input size (240x240)
56sensor.skip_frames(time=2000)     # Give the sensor time to stabilize
57
58# ------------------------------------------------------------------------------
59# Model and Overlay Loading
60# ------------------------------------------------------------------------------
61# Load the FOMO face detection model
62model = ml.Model("fomo_face_detection")  
63print("Loaded FOMO Model:", model)
64
65# Load a 1-bit PBM overlay image to obscure faces
66face_image = image.Image("/face.pbm", copy_to_fb=False)
67print("Loaded Overlay Image: face.pbm")
68
69# Convert the minimum confidence to 0..255 for find_blobs()
70threshold_list = [(math.ceil(MIN_CONFIDENCE * 255), 255)]
71
72# ------------------------------------------------------------------------------
73# FOMO Post-Processing Function
74# ------------------------------------------------------------------------------
75def fomo_post_process(model, inputs, outputs):
76    """
77    Callback function to post-process the FOMO model output.
78
79    - Converts each output channel into a grayscale image (scaled from 0..1 to 0..255).
80    - Finds blobs that exceed the threshold.
81    - Applies Non-Maximum Suppression (NMS) to refine overlapping bounding boxes.
82
83    Args:
84        model: The loaded FOMO (TensorFlow Lite) model.
85        inputs: The input image or ROI for detection.
86        outputs: The raw outputs from the model inference.
87
88    Returns:
89        A list of bounding boxes after NMS in the form:
90        [
91            [ (x1, y1, x2, y2), score ],  # for each detection of a given class
92            ...
93        ]
94        across all classes.
95    """
96    n, oh, ow, oc = model.output_shape[0]
97    
98    # NMS object to consolidate overlapping boxes
99    nms = NMS(ow, oh, inputs[0].roi)
100    
101    # Each channel corresponds to a class. For a single face class, oc=1..2, etc.
102    for i in range(oc):
103        prob_map = image.Image(outputs[0][0, :, :, i] * 255)
104        
105        # Find blobs above our threshold
106        blobs = prob_map.find_blobs(
107            threshold_list,
108            x_stride=1,
109            area_threshold=1,
110            pixels_threshold=1
111        )
112        
113        # For each blob, compute a score and store in NMS
114        for b in blobs:
115            x, y, w, h = b.rect()
116            score = prob_map.get_statistics(
117                thresholds=threshold_list,
118                roi=b.rect()
119            ).l_mean() / 255.0
120            nms.add_bounding_box(x, y, x + w, y + h, score, i)
121    
122    return nms.get_bounding_boxes()
123
124# ------------------------------------------------------------------------------
125# Ethernet Setup
126# ------------------------------------------------------------------------------
127print("- Initializing Ethernet LAN interface...")
128eth = network.LAN()     # Create a LAN object
129eth.active(True)        # Activate it
130
131# By default, we request a DHCP lease
132eth.ifconfig('dhcp')
133
134# Wait until connected
135while not eth.isconnected():
136    print("- Waiting for Ethernet connection...")
137    time.sleep_ms(1000)
138
139print("- Ethernet connected! IP information: ", eth.ifconfig())
140
141# ------------------------------------------------------------------------------
142# Create a TCP Server Socket
143# ------------------------------------------------------------------------------
144server_socket = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
145server_socket.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, True)
146server_socket.bind([HOST, PORT])
147server_socket.listen(5)
148server_socket.setblocking(True)
149
150print(f"Listening on port {PORT}...")
151
152# ------------------------------------------------------------------------------
153# Global Variables for Overlay Persistence
154# ------------------------------------------------------------------------------
155last_detections = []            # Stores the most recent bounding boxes
156last_detection_timestamp = 0     # Records the time we last saw a face
157
158# ------------------------------------------------------------------------------
159# MJPEG Streaming Function
160# ------------------------------------------------------------------------------
161def start_streaming(sock):
162    """
163    Sets up MJPEG streaming over an HTTP connection.
164
165    - Waits for a client to connect.
166    - Sends the proper multipart/x-mixed-replace HTTP headers.
167    - Continuously captures frames, performs face detection, and sends JPEG frames.
168
169    The function also manages overlay hold time. If no new detection is found,
170    but the hold time has not expired, the last detected faces remain visible.
171
172    Args:
173        sock: A listening socket that accepts incoming connections.
174    """
175    global last_detections, last_detection_timestamp
176    
177    print("Waiting for an incoming client...")
178    client, addr = sock.accept()
179    client.settimeout(5.0)
180    print(f"Client connected from {addr[0]}:{addr[1]}")
181
182    # Read the initial client request (not used), just to clear the buffer
183    client.recv(1024)
184
185    # Send HTTP headers for MJPEG
186    client.sendall(
187        "HTTP/1.1 200 OK\r\n"
188        "Server: OpenMV\r\n"
189        "Content-Type: multipart/x-mixed-replace;boundary=openmv\r\n"
190        "Cache-Control: no-cache\r\n"
191        "Pragma: no-cache\r\n\r\n"
192    )
193
194    clock = time.clock()
195
196    while True:
197        clock.tick()
198        img = sensor.snapshot()
199
200        # Run inference on the image
201        detections_by_class = model.predict([img], callback=fomo_post_process)
202        current_detections = []
203
204        # Combine detections from relevant classes (skip class 0 if background)
205        for class_idx, detection_list in enumerate(detections_by_class):
206            if class_idx == 0:
207                continue  # skip background if model includes it
208
209            # For each detection, scale and center the overlay
210            for (x, y, w, h), score in detection_list:
211                new_w = w * ENLARGEMENT_FACTOR
212                new_h = h * ENLARGEMENT_FACTOR
213
214                scale_x = new_w / face_image.width()
215                scale_y = new_h / face_image.height()
216
217                x_new = x - int((new_w - w) / 2)
218                y_new = y - int((new_h - h) / 2)
219
220                current_detections.append((x_new, y_new, scale_x, scale_y))
221
222        # If we found faces this frame, update the global info
223        if current_detections:
224            last_detections = current_detections
225            last_detection_timestamp = time.ticks_ms()
226        else:
227            # No new detections; check if hold time has expired
228            elapsed = time.ticks_ms() - last_detection_timestamp
229            if elapsed < HOLD_TIME_MS:
230                # Keep using old detections
231                current_detections = last_detections
232            else:
233                # If it's been too long, clear them
234                current_detections = []
235
236        # Draw the overlay (either fresh or held)
237        for (x_draw, y_draw, s_x, s_y) in current_detections:
238            img.draw_image(face_image, x_draw, y_draw,
239                           x_scale=s_x, y_scale=s_y)
240
241        # Convert the processed frame to JPEG for streaming
242        cframe = img.to_jpeg(quality=35, copy=True)
243        header = (
244            "\r\n--openmv\r\n"
245            "Content-Type: image/jpeg\r\n"
246            f"Content-Length:{cframe.size()}\r\n\r\n"
247        )
248
249        # Send the multipart image to the client
250        client.sendall(header)
251        client.sendall(cframe)
252
253        # Print FPS to the console (optional)
254        print(f"FPS: {clock.fps()}")
255
256# ------------------------------------------------------------------------------
257# Main Loop
258# ------------------------------------------------------------------------------
259while True:
260    try:
261        # Start streaming on our server socket
262        start_streaming(server_socket)
263    except OSError as e:
264        # If client disconnects or other socket errors, go back and wait again
265        print("Socket error:", e)

"""
Automatic Face Ban (Ethernet Version)
Name: portenta_faceban_ethernet.py
Purpose:
    This script demonstrates how to automatically censor (ban) faces in
    a live video stream for privacy protection using an Ethernet connection.
    Using TensorFlow Lite FOMO (Fast Object Detection Model) on a 
    Portenta H7 + Portenta Vision Shield, faces are detected in real time 
    and an overlay is used to obscure them before streaming via MJPEG over HTTP.
    
    By enlarging the overlay and keeping it on-screen for a short hold time
    (even if the face momentarily disappears), this solution helps ensure
    that no identifiable faces are exposed, which is crucial for compliance
    with privacy regulations in environments like schools, hospitals,
    or public buildings.

Author: Arduino Product Experience Team
Version: 1.0 (27/01/25)

Key Features:
    - MJPEG video streaming over HTTP for real-time viewing.
    - Automatic face detection and overlay (face ban/censorship).
    - Configurable hold time to keep obscuring the face after detection loss.
    - Ideal for IoT-edge deployments where on-device processing is required.
"""

import sensor
import time
import image
import network
import socket
import ml
from ml.utils import NMS
import math

# ------------------------------------------------------------------------------
# User Configuration
# ------------------------------------------------------------------------------
HOST = ""  # Binds to the first available interface
PORT = 8080

# ------------------------------------------------------------------------------
# Detection Parameters
# ------------------------------------------------------------------------------
MIN_CONFIDENCE = 0.4           # Minimum confidence threshold for faces
ENLARGEMENT_FACTOR = 3.5       # Scale factor to enlarge the overlay over the face
HOLD_TIME_MS = 1500            # How long (ms) to keep overlay after losing detection

# ------------------------------------------------------------------------------
# Camera Setup
# ------------------------------------------------------------------------------
sensor.reset()
sensor.set_pixformat(sensor.GRAYSCALE)
sensor.set_framesize(sensor.QVGA)
sensor.set_windowing((240, 240))  # Typical FOMO model input size (240x240)
sensor.skip_frames(time=2000)     # Give the sensor time to stabilize

# ------------------------------------------------------------------------------
# Model and Overlay Loading
# ------------------------------------------------------------------------------
# Load the FOMO face detection model
model = ml.Model("fomo_face_detection")  
print("Loaded FOMO Model:", model)

# Load a 1-bit PBM overlay image to obscure faces
face_image = image.Image("/face.pbm", copy_to_fb=False)
print("Loaded Overlay Image: face.pbm")

# Convert the minimum confidence to 0..255 for find_blobs()
threshold_list = [(math.ceil(MIN_CONFIDENCE * 255), 255)]

# ------------------------------------------------------------------------------
# FOMO Post-Processing Function
# ------------------------------------------------------------------------------
def fomo_post_process(model, inputs, outputs):
    """
    Callback function to post-process the FOMO model output.

    - Converts each output channel into a grayscale image (scaled from 0..1 to 0..255).
    - Finds blobs that exceed the threshold.
    - Applies Non-Maximum Suppression (NMS) to refine overlapping bounding boxes.

    Args:
        model: The loaded FOMO (TensorFlow Lite) model.
        inputs: The input image or ROI for detection.
        outputs: The raw outputs from the model inference.

    Returns:
        A list of bounding boxes after NMS in the form:
        [
            [ (x1, y1, x2, y2), score ],  # for each detection of a given class
            ...
        ]
        across all classes.
    """
    n, oh, ow, oc = model.output_shape[0]
    
    # NMS object to consolidate overlapping boxes
    nms = NMS(ow, oh, inputs[0].roi)
    
    # Each channel corresponds to a class. For a single face class, oc=1..2, etc.
    for i in range(oc):
        prob_map = image.Image(outputs[0][0, :, :, i] * 255)
        
        # Find blobs above our threshold
        blobs = prob_map.find_blobs(
            threshold_list,
            x_stride=1,
            area_threshold=1,
            pixels_threshold=1
        )
        
        # For each blob, compute a score and store in NMS
        for b in blobs:
            x, y, w, h = b.rect()
            score = prob_map.get_statistics(
                thresholds=threshold_list,
                roi=b.rect()
            ).l_mean() / 255.0
            nms.add_bounding_box(x, y, x + w, y + h, score, i)
    
    return nms.get_bounding_boxes()

# ------------------------------------------------------------------------------
# Ethernet Setup
# ------------------------------------------------------------------------------
print("- Initializing Ethernet LAN interface...")
eth = network.LAN()     # Create a LAN object
eth.active(True)        # Activate it

# By default, we request a DHCP lease
eth.ifconfig('dhcp')

# Wait until connected
while not eth.isconnected():
    print("- Waiting for Ethernet connection...")
    time.sleep_ms(1000)

print("- Ethernet connected! IP information: ", eth.ifconfig())

# ------------------------------------------------------------------------------
# Create a TCP Server Socket
# ------------------------------------------------------------------------------
server_socket = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
server_socket.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, True)
server_socket.bind([HOST, PORT])
server_socket.listen(5)
server_socket.setblocking(True)

print(f"Listening on port {PORT}...")

# ------------------------------------------------------------------------------
# Global Variables for Overlay Persistence
# ------------------------------------------------------------------------------
last_detections = []            # Stores the most recent bounding boxes
last_detection_timestamp = 0     # Records the time we last saw a face

# ------------------------------------------------------------------------------
# MJPEG Streaming Function
# ------------------------------------------------------------------------------
def start_streaming(sock):
    """
    Sets up MJPEG streaming over an HTTP connection.

    - Waits for a client to connect.
    - Sends the proper multipart/x-mixed-replace HTTP headers.
    - Continuously captures frames, performs face detection, and sends JPEG frames.

    The function also manages overlay hold time. If no new detection is found,
    but the hold time has not expired, the last detected faces remain visible.

    Args:
        sock: A listening socket that accepts incoming connections.
    """
    global last_detections, last_detection_timestamp
    
    print("Waiting for an incoming client...")
    client, addr = sock.accept()
    client.settimeout(5.0)
    print(f"Client connected from {addr[0]}:{addr[1]}")

    # Read the initial client request (not used), just to clear the buffer
    client.recv(1024)

    # Send HTTP headers for MJPEG
    client.sendall(
        "HTTP/1.1 200 OK\r\n"
        "Server: OpenMV\r\n"
        "Content-Type: multipart/x-mixed-replace;boundary=openmv\r\n"
        "Cache-Control: no-cache\r\n"
        "Pragma: no-cache\r\n\r\n"
    )

    clock = time.clock()

    while True:
        clock.tick()
        img = sensor.snapshot()

        # Run inference on the image
        detections_by_class = model.predict([img], callback=fomo_post_process)
        current_detections = []

        # Combine detections from relevant classes (skip class 0 if background)
        for class_idx, detection_list in enumerate(detections_by_class):
            if class_idx == 0:
                continue  # skip background if model includes it

            # For each detection, scale and center the overlay
            for (x, y, w, h), score in detection_list:
                new_w = w * ENLARGEMENT_FACTOR
                new_h = h * ENLARGEMENT_FACTOR

                scale_x = new_w / face_image.width()
                scale_y = new_h / face_image.height()

                x_new = x - int((new_w - w) / 2)
                y_new = y - int((new_h - h) / 2)

                current_detections.append((x_new, y_new, scale_x, scale_y))

        # If we found faces this frame, update the global info
        if current_detections:
            last_detections = current_detections
            last_detection_timestamp = time.ticks_ms()
        else:
            # No new detections; check if hold time has expired
            elapsed = time.ticks_ms() - last_detection_timestamp
            if elapsed < HOLD_TIME_MS:
                # Keep using old detections
                current_detections = last_detections
            else:
                # If it's been too long, clear them
                current_detections = []

        # Draw the overlay (either fresh or held)
        for (x_draw, y_draw, s_x, s_y) in current_detections:
            img.draw_image(face_image, x_draw, y_draw,
                           x_scale=s_x, y_scale=s_y)

        # Convert the processed frame to JPEG for streaming
        cframe = img.to_jpeg(quality=35, copy=True)
        header = (
            "\r\n--openmv\r\n"
            "Content-Type: image/jpeg\r\n"
            f"Content-Length:{cframe.size()}\r\n\r\n"
        )

        # Send the multipart image to the client
        client.sendall(header)
        client.sendall(cframe)

        # Print FPS to the console (optional)
        print(f"FPS: {clock.fps()}")

# ------------------------------------------------------------------------------
# Main Loop
# ------------------------------------------------------------------------------
while True:
    try:
        # Start streaming on our server socket
        start_streaming(server_socket)
    except OSError as e:
        # If client disconnects or other socket errors, go back and wait again
        print("Socket error:", e)

Face Ban (Wi-Fi)

python

The Wi-Fi version of the face ban script.

1"""
2Automatic Face Ban (Wi-Fi Version)
3Name: portenta_faceban_wifi.py
4Purpose:
5    This script demonstrates how to automatically censor (ban) faces in
6    a live video stream for privacy protection using an Wi-Fi connection.
7    Using TensorFlow Lite FOMO (Fast Object Detection Model) on a 
8    Portenta H7 + Portenta Vision Shield, faces are detected in real time 
9    and an overlay is used to obscure them before streaming via MJPEG over HTTP.
10
11    By enlarging the overlay and keeping it on-screen for a short hold time
12    (even if the face momentarily disappears), this solution helps ensure
13    that no identifiable faces are exposed, which is crucial for compliance
14    with privacy regulations in environments like schools, hospitals,
15    or public buildings.
16
17Author: Arduino Product Experience Team
18Version: 1.0 (27/01/25)
19
20Key Features:
21    - MJPEG video streaming over HTTP for real-time viewing.
22    - Automatic face detection and overlay (face ban/censorship).
23    - Configurable hold time to keep obscuring the face after detection loss.
24    - Ideal for IoT-edge deployments where on-device processing is required.
25"""
26
27import sensor
28import time
29import image
30import network
31import socket
32import ml
33from ml.utils import NMS
34import math
35
36# ------------------------------------------------------------------------------
37# User Configuration
38# ------------------------------------------------------------------------------
39SSID = "YOUR_SSID"            # Wi-Fi SSID
40KEY =  "YOUR_KEY"             # Wi-Fi Password
41HOST = ""                     # Use the first available interface
42PORT = 8080                   # TCP Port for the HTTP server
43
44# ------------------------------------------------------------------------------
45# Detection Parameters
46# ------------------------------------------------------------------------------
47MIN_CONFIDENCE = 0.4          # Minimum confidence threshold for detections
48ENLARGEMENT_FACTOR = 3.5      # Scale factor for the overlay relative to the detected face
49HOLD_TIME_MS = 1500           # Time (ms) to keep overlay after last detection
50
51# ------------------------------------------------------------------------------
52# Camera Setup
53# ------------------------------------------------------------------------------
54sensor.reset()
55sensor.set_pixformat(sensor.GRAYSCALE)
56sensor.set_framesize(sensor.QVGA)
57sensor.set_windowing((240, 240))  # Common FOMO model input size
58sensor.skip_frames(time=2000)     # Wait for the sensor to stabilize
59
60# ------------------------------------------------------------------------------
61# Model and Overlay Loading
62# ------------------------------------------------------------------------------
63# Load the FOMO face detection model
64model = ml.Model("fomo_face_detection")
65print("- Loaded FOMO model:", model)
66
67# Load the overlay image (1-bit PBM)
68face_image = image.Image("/face.pbm", copy_to_fb=False)
69print("- Loaded overlay image: face.pbm")
70
71# Prepare threshold based on minimum confidence
72threshold_list = [(math.ceil(MIN_CONFIDENCE * 255), 255)]
73
74# ------------------------------------------------------------------------------
75# FOMO Post-processing Function
76# ------------------------------------------------------------------------------
77def fomo_post_process(model, inputs, outputs):
78    """
79    Callback function to post-process the FOMO model output.
80
81    - Converts each output channel into a grayscale image (scaled from 0..1 to 0..255).
82    - Finds blobs that exceed the threshold.
83    - Applies Non-Maximum Suppression (NMS) to refine overlapping bounding boxes.
84
85    Args:
86        model: The loaded FOMO (TensorFlow Lite) model.
87        inputs: The input image or ROI for detection.
88        outputs: The raw outputs from the model inference.
89
90    Returns:
91        A list of bounding boxes after NMS in the form:
92        [
93            [ (x1, y1, x2, y2), score ],  # for each detection of a given class
94            ...
95        ]
96        across all classes.
97    """
98    n, oh, ow, oc = model.output_shape[0]
99
100    # Instantiate an NMS object with the output dims and ROI
101    nms = NMS(ow, oh, inputs[0].roi)
102
103    # For each class channel, detect blobs
104    for i in range(oc):
105        # Scale model output from [0.0..1.0] to [0..255]
106        prob_map = image.Image(outputs[0][0, :, :, i] * 255)
107
108        # Locate blobs above the threshold
109        blobs = prob_map.find_blobs(
110            threshold_list,
111            x_stride=1,
112            area_threshold=1,
113            pixels_threshold=1
114        )
115
116        # For each blob, calculate a score and store the bounding box
117        for b in blobs:
118            x, y, w, h = b.rect()
119            score = prob_map.get_statistics(
120                thresholds=threshold_list,
121                roi=b.rect()
122            ).l_mean() / 255.0
123            nms.add_bounding_box(x, y, x + w, y + h, score, i)
124
125    return nms.get_bounding_boxes()
126
127# ------------------------------------------------------------------------------
128# Wi-Fi Setup
129# ------------------------------------------------------------------------------
130wlan = network.WLAN(network.STA_IF)
131wlan.active(True)
132wlan.connect(SSID, KEY)
133
134while not wlan.isconnected():
135    print(f"- Trying to connect to \"{SSID}\"...")
136    time.sleep_ms(1000)
137
138print("- Wi-Fi connected! IP information:", wlan.ifconfig())
139
140# ------------------------------------------------------------------------------
141# Socket Server Setup
142# ------------------------------------------------------------------------------
143server_socket = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
144server_socket.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, True)
145server_socket.bind([HOST, PORT])
146server_socket.listen(5)
147server_socket.setblocking(True)
148
149# ------------------------------------------------------------------------------
150# Global Variables for Holding Detections
151# ------------------------------------------------------------------------------
152last_detections = []
153last_detection_timestamp = 0
154
155# ------------------------------------------------------------------------------
156# Streaming Function
157# ------------------------------------------------------------------------------
158def start_streaming(sock):
159    """
160    Sets up MJPEG streaming over an HTTP connection.
161
162    - Waits for a client to connect.
163    - Sends the proper multipart/x-mixed-replace HTTP headers.
164    - Continuously captures frames, performs face detection, and sends JPEG frames.
165
166    The function also manages overlay hold time. If no new detection is found,
167    but the hold time has not expired, the last detected faces remain visible.
168
169    Args:
170        sock: A listening socket that accepts incoming connections.
171    """
172    global last_detections, last_detection_timestamp
173
174    print("- Waiting for connections...")
175    client, addr = sock.accept()
176    client.settimeout(5.0)
177    print(f"- Connected to {addr[0]}:{addr[1]}")
178
179    # Read initial client request (not used) to clear the buffer
180    client.recv(1024)
181
182    # Send HTTP headers for MJPEG streaming
183    client.sendall(
184        "HTTP/1.1 200 OK\r\n"
185        "Server: OpenMV\r\n"
186        "Content-Type: multipart/x-mixed-replace;boundary=openmv\r\n"
187        "Cache-Control: no-cache\r\n"
188        "Pragma: no-cache\r\n\r\n"
189    )
190
191    # Create a clock to measure FPS (optional)
192    clock = time.clock()
193
194    while True:
195        clock.tick()
196        img = sensor.snapshot()
197
198        # Perform face detection via the FOMO model
199        detections_by_class = model.predict([img], callback=fomo_post_process)
200        current_detections = []
201
202        # Gather detections from classes > 0 (assuming 0 is background)
203        for class_idx, detection_list in enumerate(detections_by_class):
204            if class_idx == 0:
205                continue  # skip background
206            for (x, y, w, h), score in detection_list:
207                # Calculate enlarged overlay dimensions
208                new_w = w * ENLARGEMENT_FACTOR
209                new_h = h * ENLARGEMENT_FACTOR
210
211                # Determine scaling factors for face_image
212                scale_x = new_w / face_image.width()
213                scale_y = new_h / face_image.height()
214
215                # Shift x,y so the overlay remains centered on the face
216                x_new = x - int((new_w - w) / 2)
217                y_new = y - int((new_h - h) / 2)
218
219                # Store detection info for drawing
220                current_detections.append((x_new, y_new, scale_x, scale_y))
221
222        # Update or retain previous detections
223        if current_detections:
224            last_detections = current_detections
225            last_detection_timestamp = time.ticks_ms()
226        else:
227            # If no new detections, check if we should hold the previous ones
228            elapsed = time.ticks_ms() - last_detection_timestamp
229            if elapsed < HOLD_TIME_MS:
230                current_detections = last_detections
231            else:
232                current_detections = []
233
234        # Draw overlays (fresh or held) on the image
235        for (x_draw, y_draw, s_x, s_y) in current_detections:
236            img.draw_image(face_image, x_draw, y_draw,
237                           x_scale=s_x,
238                           y_scale=s_y)
239
240        # Convert the final frame to JPEG
241        cframe = img.to_jpeg(quality=35, copy=True)
242        header = (
243            "\r\n--openmv\r\n"
244            "Content-Type: image/jpeg\r\n"
245            f"Content-Length:{cframe.size()}\r\n\r\n"
246        )
247
248        # Send the multipart frame
249        client.sendall(header)
250        client.sendall(cframe)
251
252        # Debug: Print FPS
253        print(f"- FPS: {clock.fps()}")
254
255# ------------------------------------------------------------------------------
256# Main Loop
257# ------------------------------------------------------------------------------
258while True:
259    try:
260        start_streaming(server_socket)
261    except OSError as e:
262        # If the connection fails or is dropped, try again
263        print("- Socket error: ", e)

"""
Automatic Face Ban (Wi-Fi Version)
Name: portenta_faceban_wifi.py
Purpose:
    This script demonstrates how to automatically censor (ban) faces in
    a live video stream for privacy protection using an Wi-Fi connection.
    Using TensorFlow Lite FOMO (Fast Object Detection Model) on a 
    Portenta H7 + Portenta Vision Shield, faces are detected in real time 
    and an overlay is used to obscure them before streaming via MJPEG over HTTP.

    By enlarging the overlay and keeping it on-screen for a short hold time
    (even if the face momentarily disappears), this solution helps ensure
    that no identifiable faces are exposed, which is crucial for compliance
    with privacy regulations in environments like schools, hospitals,
    or public buildings.

Author: Arduino Product Experience Team
Version: 1.0 (27/01/25)

Key Features:
    - MJPEG video streaming over HTTP for real-time viewing.
    - Automatic face detection and overlay (face ban/censorship).
    - Configurable hold time to keep obscuring the face after detection loss.
    - Ideal for IoT-edge deployments where on-device processing is required.
"""

import sensor
import time
import image
import network
import socket
import ml
from ml.utils import NMS
import math

# ------------------------------------------------------------------------------
# User Configuration
# ------------------------------------------------------------------------------
SSID = "YOUR_SSID"            # Wi-Fi SSID
KEY =  "YOUR_KEY"             # Wi-Fi Password
HOST = ""                     # Use the first available interface
PORT = 8080                   # TCP Port for the HTTP server

# ------------------------------------------------------------------------------
# Detection Parameters
# ------------------------------------------------------------------------------
MIN_CONFIDENCE = 0.4          # Minimum confidence threshold for detections
ENLARGEMENT_FACTOR = 3.5      # Scale factor for the overlay relative to the detected face
HOLD_TIME_MS = 1500           # Time (ms) to keep overlay after last detection

# ------------------------------------------------------------------------------
# Camera Setup
# ------------------------------------------------------------------------------
sensor.reset()
sensor.set_pixformat(sensor.GRAYSCALE)
sensor.set_framesize(sensor.QVGA)
sensor.set_windowing((240, 240))  # Common FOMO model input size
sensor.skip_frames(time=2000)     # Wait for the sensor to stabilize

# ------------------------------------------------------------------------------
# Model and Overlay Loading
# ------------------------------------------------------------------------------
# Load the FOMO face detection model
model = ml.Model("fomo_face_detection")
print("- Loaded FOMO model:", model)

# Load the overlay image (1-bit PBM)
face_image = image.Image("/face.pbm", copy_to_fb=False)
print("- Loaded overlay image: face.pbm")

# Prepare threshold based on minimum confidence
threshold_list = [(math.ceil(MIN_CONFIDENCE * 255), 255)]

# ------------------------------------------------------------------------------
# FOMO Post-processing Function
# ------------------------------------------------------------------------------
def fomo_post_process(model, inputs, outputs):
    """
    Callback function to post-process the FOMO model output.

    - Converts each output channel into a grayscale image (scaled from 0..1 to 0..255).
    - Finds blobs that exceed the threshold.
    - Applies Non-Maximum Suppression (NMS) to refine overlapping bounding boxes.

    Args:
        model: The loaded FOMO (TensorFlow Lite) model.
        inputs: The input image or ROI for detection.
        outputs: The raw outputs from the model inference.

    Returns:
        A list of bounding boxes after NMS in the form:
        [
            [ (x1, y1, x2, y2), score ],  # for each detection of a given class
            ...
        ]
        across all classes.
    """
    n, oh, ow, oc = model.output_shape[0]

    # Instantiate an NMS object with the output dims and ROI
    nms = NMS(ow, oh, inputs[0].roi)

    # For each class channel, detect blobs
    for i in range(oc):
        # Scale model output from [0.0..1.0] to [0..255]
        prob_map = image.Image(outputs[0][0, :, :, i] * 255)

        # Locate blobs above the threshold
        blobs = prob_map.find_blobs(
            threshold_list,
            x_stride=1,
            area_threshold=1,
            pixels_threshold=1
        )

        # For each blob, calculate a score and store the bounding box
        for b in blobs:
            x, y, w, h = b.rect()
            score = prob_map.get_statistics(
                thresholds=threshold_list,
                roi=b.rect()
            ).l_mean() / 255.0
            nms.add_bounding_box(x, y, x + w, y + h, score, i)

    return nms.get_bounding_boxes()

# ------------------------------------------------------------------------------
# Wi-Fi Setup
# ------------------------------------------------------------------------------
wlan = network.WLAN(network.STA_IF)
wlan.active(True)
wlan.connect(SSID, KEY)

while not wlan.isconnected():
    print(f"- Trying to connect to \"{SSID}\"...")
    time.sleep_ms(1000)

print("- Wi-Fi connected! IP information:", wlan.ifconfig())

# ------------------------------------------------------------------------------
# Socket Server Setup
# ------------------------------------------------------------------------------
server_socket = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
server_socket.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, True)
server_socket.bind([HOST, PORT])
server_socket.listen(5)
server_socket.setblocking(True)

# ------------------------------------------------------------------------------
# Global Variables for Holding Detections
# ------------------------------------------------------------------------------
last_detections = []
last_detection_timestamp = 0

# ------------------------------------------------------------------------------
# Streaming Function
# ------------------------------------------------------------------------------
def start_streaming(sock):
    """
    Sets up MJPEG streaming over an HTTP connection.

    - Waits for a client to connect.
    - Sends the proper multipart/x-mixed-replace HTTP headers.
    - Continuously captures frames, performs face detection, and sends JPEG frames.

    The function also manages overlay hold time. If no new detection is found,
    but the hold time has not expired, the last detected faces remain visible.

    Args:
        sock: A listening socket that accepts incoming connections.
    """
    global last_detections, last_detection_timestamp

    print("- Waiting for connections...")
    client, addr = sock.accept()
    client.settimeout(5.0)
    print(f"- Connected to {addr[0]}:{addr[1]}")

    # Read initial client request (not used) to clear the buffer
    client.recv(1024)

    # Send HTTP headers for MJPEG streaming
    client.sendall(
        "HTTP/1.1 200 OK\r\n"
        "Server: OpenMV\r\n"
        "Content-Type: multipart/x-mixed-replace;boundary=openmv\r\n"
        "Cache-Control: no-cache\r\n"
        "Pragma: no-cache\r\n\r\n"
    )

    # Create a clock to measure FPS (optional)
    clock = time.clock()

    while True:
        clock.tick()
        img = sensor.snapshot()

        # Perform face detection via the FOMO model
        detections_by_class = model.predict([img], callback=fomo_post_process)
        current_detections = []

        # Gather detections from classes > 0 (assuming 0 is background)
        for class_idx, detection_list in enumerate(detections_by_class):
            if class_idx == 0:
                continue  # skip background
            for (x, y, w, h), score in detection_list:
                # Calculate enlarged overlay dimensions
                new_w = w * ENLARGEMENT_FACTOR
                new_h = h * ENLARGEMENT_FACTOR

                # Determine scaling factors for face_image
                scale_x = new_w / face_image.width()
                scale_y = new_h / face_image.height()

                # Shift x,y so the overlay remains centered on the face
                x_new = x - int((new_w - w) / 2)
                y_new = y - int((new_h - h) / 2)

                # Store detection info for drawing
                current_detections.append((x_new, y_new, scale_x, scale_y))

        # Update or retain previous detections
        if current_detections:
            last_detections = current_detections
            last_detection_timestamp = time.ticks_ms()
        else:
            # If no new detections, check if we should hold the previous ones
            elapsed = time.ticks_ms() - last_detection_timestamp
            if elapsed < HOLD_TIME_MS:
                current_detections = last_detections
            else:
                current_detections = []

        # Draw overlays (fresh or held) on the image
        for (x_draw, y_draw, s_x, s_y) in current_detections:
            img.draw_image(face_image, x_draw, y_draw,
                           x_scale=s_x,
                           y_scale=s_y)

        # Convert the final frame to JPEG
        cframe = img.to_jpeg(quality=35, copy=True)
        header = (
            "\r\n--openmv\r\n"
            "Content-Type: image/jpeg\r\n"
            f"Content-Length:{cframe.size()}\r\n\r\n"
        )

        # Send the multipart frame
        client.sendall(header)
        client.sendall(cframe)

        # Debug: Print FPS
        print(f"- FPS: {clock.fps()}")

# ------------------------------------------------------------------------------
# Main Loop
# ------------------------------------------------------------------------------
while True:
    try:
        start_streaming(server_socket)
    except OSError as e:
        # If the connection fails or is dropped, try again
        print("- Socket error: ", e)

Downloadable files

face.pbm

This is the image used for overlay.

face.pbm

Documentation

Portenta Vision Shield User Manual

Learn how to set up, configure and use the shield.

https://docs.arduino.cc/tutorials/portenta-vision-shield/user-manual

Comments

Only logged in users can leave comments

jabagur

0 Followers

•