Face Recognition on Arduino VENTUNO Q

ROS 2 Jazzy

Arduino®️ VENTUNO™ Q

<html><body><figure><img src="https://projects.arduinocontent.cc/7ca392e7-0e01-491e-a342-84eee61389ea.png"/><figcaption>Pipeline architecture diagram</figcaption></figure>sample_face_recognition answers the harder question that usually comes next: "whose face is this?" It's a C++ ROS 2 node built entirely on open-source OpenCV Zoo models — no proprietary detector, no cloud lookup — that detects faces with the YuNet model and identifies them with the SFace embedding model, matching against a small local database you build yourself.<h3>Prerequisites</h3><figure><img src="https://projects.arduinocontent.cc/68d19a3b-ec95-4c45-9d7f-064d1c48eecc.png"/><figcaption>Prerequisites table</figcaption></figure><h3>What the sample does</h3>The node is organized as four cooperating submodules inside a single composable node, each with a narrow job:<figure><img src="https://projects.arduinocontent.cc/05059c11-91a5-48f1-b402-2239d098b3f0.png"/><figcaption>Topics table</figcaption></figure><figure><img src="https://projects.arduinocontent.cc/9dafea95-1266-4efe-b230-101202d26125.png"/><figcaption>Submodules table</figcaption></figure><h3>Why the models need a fixed, "simplified" input shape</h3>YuNet and SFace, as published in opencv_zoo, ship with dynamic input shapes so they can run at whatever resolution a caller provides. A ROS 2 node that has to allocate fixed-size tensors up front doesn't want that flexibility — it wants one known shape it can commit to at startup. That's exactly what the onnxsim ... --overwrite-input-shape step in the build does: it locks the detector to a 1×3×320×320 input and the recognizer to 1×3×112×112, producing the _static.onnx variants the launch file actually loads. Skip this step and the node either fails to load the model or silently falls back to shapes the rest of the pipeline wasn't built to expect.<h3>How matching actually works</h3>SFace doesn't output a name — it outputs a fixed-length embedding vector for a detected face. Recognition happens by comparing that vector against every embedding already stored in the face database and accepting the closest match only if its similarity score clears similar_threshold (default 0.363). Raise the threshold to make the system more conservative about claiming a match (fewer false positives, more "unknown" results); lower it if genuine matches are being rejected because of lighting or angle differences between the registered image and the live frame. The very first time the node runs, it also has to build this database from scratch from whatever images sit under image_data_path — watch the startup log for detect people: &lt;name&gt; lines and the generated face_features.bin / face_names.txt files as confirmation it worked.<h3>Try it yourself, step by step</h3>Step 1 — Download and reshape the models<pre>mkdir -p ~/model &amp;&amp; cd ~/model </pre><pre>wget "https://media.githubusercontent.com/media/opencv/opencv_zoo/42802fb13aecbf59542f2cb1125a6306da882bf2/models/face_detection_yunet/face_detection_yunet_2021dec.onnx?download=true" -O face_detection_yunet_2021dec.onnx </pre><pre>wget "https://media.githubusercontent.com/media/opencv/opencv_zoo/refs/heads/main/models/face_recognition_sface/face_recognition_sface_2021dec.onnx?download=true" -O face_recognition_sface_2021dec.onnx </pre><pre> </pre><pre>python3 -m onnxsim ./face_detection_yunet_2021dec.onnx ./face_detection_yunet_2021dec_static.onnx \ </pre><pre>	 --overwrite-input-shape input:1,3,320,320 </pre><pre>python3 -m onnxsim ./face_recognition_sface_2021dec.onnx ./face_recognition_sface_2021dec_static.onnx \ </pre><pre>	 --overwrite-input-shape data:1,3,112,112 </pre>Expected result: two new _static.onnx files appear alongside the originals — these, not the dynamic-shape originals, are what the launch file loads.Step 2 — Build the workspace<pre>mkdir -p ~/qrb_ros_sample_ws/src &amp;&amp; cd ~/qrb_ros_sample_ws/src </pre><pre>git clone https://github.com/qualcomm-qrb-ros/qrb_ros_samples.git </pre><pre>cd qrb_ros_samples/ai_vision/sample_face_recognition/ </pre><pre>mkdir model &amp;&amp; cp -vf ~/model/* ./model/ </pre><pre> </pre><pre>source /opt/ros/jazzy/setup.bash </pre><pre>colcon build --packages-up-to sample_face_recognition </pre>Expected result: sample_face_recognition and its dependency sample_face_customed_msgs build without errors.Step 3 — Launch the recognition node<pre>source /opt/ros/jazzy/setup.bash </pre><pre>export ROS_DOMAIN_ID=66 </pre><pre>source install/setup.bash </pre><pre>ros2 launch sample_face_recognition face_recognition.launch.py </pre>Expected result: the log shows both models loading, the face database being built from the bundled sample images (log lines like detect people: old_man), and start QueryName service — the node is now waiting for frames on /input_faces.Step 4 — Feed it a test image<pre># in a second terminal, same ROS_DOMAIN_ID </pre><pre>source /opt/ros/jazzy/setup.bash </pre><pre>export ROS_DOMAIN_ID=66 </pre><pre>cd qrb_ros_samples/ai_vision/sample_face_recognition/ </pre><pre>source install/setup.bash </pre><pre> </pre><pre>python3 ./sample_face_recognition/scripts/image_pub.py \ </pre><pre>	 -i ./resource/test_img.jpg -t /input_faces -r 1 --frame-id cam0 --encoding bgr8 </pre>Expected result: the first terminal starts logging detecting N &lt;name&gt; score=... size=... lines for every face found in the test image, matched (or not) against the database.Step 5 — View the annotated result<pre>ros2 run rqt_image_view rqt_image_view /detected_faces_overlay </pre>Expected result: the test image with each detected face boxed and labeled with either a registered name or an "unknown" indicator, depending on whether it cleared similar_threshold.<h3>Troubleshooting</h3><figure><img src="https://projects.arduinocontent.cc/5af993c0-08c8-4ca2-bb1a-c25f71155849.png"/><figcaption>Troubleshooting table</figcaption></figure><h3>Detection vs. recognition, and where this fits</h3>It's worth keeping the two sibling samples straight: sample_face_detection is the fast, HTP-accelerated stage that answers "where are the faces and their landmarks" at high frame rate; sample_face_recognition is the heavier, identity-matching stage that answers "whose face is this," at the cost of maintaining a database and tuning a similarity threshold. In a larger system, it's entirely reasonable to run detection continuously and only invoke recognition — or a model like this one — when a downstream task genuinely needs identity, rather than on every frame.</body></html>

Face Recognition on Arduino VENTUNO Q

Real-time face detection (NPU via QNN) plus face recognition (CPU, SFace ONNX) on Arduino VENTUNO Q.