Give Your Treadmill Eyes with Edge AI

Arduino®️ VENTUNO™ Q

Qualcomm AI Runtime (QAIRT)

ONNX Runtime

C960 Web Camera

<html><body><h3>AI Treadmill: Turn a Regular Treadmill into a Bluetooth Smart Treadmill</h3>Full blog and detailed reproduction notes: https://github.com/munoz0raul/treadmill/blob/main/docs/PROJECT_HUB.md<figure><img src="https://projects.arduinocontent.cc/ece176f0-04af-454c-9805-36f2a1a70409.jpg"/><figcaption></figcaption></figure><h3>Introduction</h3>I work from home, and getting out of the house to exercise has become harder than it should be. After a few doctors reminded me that I needed to take exercise more seriously, my wife rented a treadmill. That helped, but only a little. The treadmill was useful, but walking in place was still boring. After a few weeks, my 3D printer was still more interesting than spending an hour on the treadmill. So I tried to make the treadmill more fun. I had seen people connect bikes and treadmills to virtual training games, and I wondered: could I do the same without modifying the treadmill? That constraint mattered a lot. The treadmill was rented. I could not open it, drill it, add magnets, wire into the motor controller, or attach anything permanent. I needed something completely external. The idea became this project: use a camera and local AI to watch the treadmill, estimate the walking speed, and broadcast that speed over Bluetooth as if the treadmill were smart. <h3>What this project does</h3>AI Treadmill turns a regular belt treadmill into a virtual smart treadmill without modifying the treadmill. A USB camera looks at the treadmill from the side. An Arduino VENTUNO Q runs a local video AI model that estimates treadmill speed from a rolling one-second video clip. The board then advertises itself over Bluetooth Low Energy using FTMS, the Bluetooth Fitness Machine Service. FTMS is the standard Bluetooth profile used by many fitness apps and games to talk to treadmills, bikes, rowers, and other fitness machines. By speaking FTMS, the VENTUNO Q can appear to compatible apps as a normal smart treadmill. The flow is: <pre>	rented treadmill </pre><pre>-&gt; USB camera </pre><pre>-&gt; local CNN speed estimator </pre><pre>-&gt; Bluetooth FTMS </pre><pre>-&gt; fitness game or app </pre>No treadmill modification. No magnets. No optical encoder. No belt sensor. Just a camera, local AI, and a standard Bluetooth fitness profile. <h3>Why local AI?</h3>A camera pointed at a treadmill is still a camera pointed inside my home. I did not want images or video going to the Internet just to estimate speed. This project keeps the video local. The camera frames stay on the VENTUNO Q, inference runs on the board, and only the final speed value is sent over Bluetooth. <h3>What you will build</h3>By the end of the project, you will have: <ol><li>A web app that shows a live camera feed and estimated treadmill speed. </li><li>A local video model running on the VENTUNO Q. </li><li>A Bluetooth FTMS treadmill peripheral named AI Treadmill . </li><li>A way for compatible fitness apps and games to read speed from a non-smart treadmill. </li><li>A reproducible dataset and training path if you want to rebuild or adapt the model. </li><li>An optional NPU acceleration path using Qualcomm AI Runtime. </li></ol><h3>Software</h3>The default runtime path uses: <ol><li>Python 3 </li><li>OpenCV </li><li>NumPy </li><li>ONNX Runtime </li><li>BlueZ Bluetooth stack </li><li>bless for BLE peripheral support </li><li>bleak for the optional FTMS diagnostic probe </li></ol>The optional training path uses: <ol><li>PyTorch </li><li>torchvision </li><li>Kinetics-400 pretrained <code class="code">mc3_18 </code>video backbone </li></ol>The optional NPU path uses: <ol><li>Qualcomm AI Runtime SDK, QAIRT </li><li>A16W8 quantization </li><li>Hexagon HTP V75 context binary </li><li><code class="code">qnn-net-run </code></li></ol><h3>Camera placement</h3>Camera placement matters more than camera quality. Use a side view: <ol><li>Put the webcam on the side of the treadmill. </li><li>Keep the belt and the walker visible. </li><li>Use stable lighting.Keep the camera fixed. </li><li>Avoid moving background objects. </li></ol> <figure><img src="https://projects.arduinocontent.cc/7b36c41b-f2a9-4c53-81bb-7459813ca47c.jpg"/><figcaption></figcaption></figure> The training and live inference code use a central crop of the camera image. If you change the camera angle or framing, make sure the crop used for training and live inference matches. <h3>How to read the commands</h3>Commands in this tutorial run in one of two places: <ol><li>VENTUNO Q : the board connected to the USB camera and Bluetooth. </li><li>HOST machine : your workstation, used only for training or the optional NPU build. </li></ol>For the main runtime path, you only need the VENTUNO Q. <h3>VENTUNO Q setup</h3>Run on: VENTUNO Q Clone the repository: <pre>git clone https://github.com/munoz0raul/treadmill.git treadmill </pre><pre>cd treadmill </pre><pre>export REPO="$PWD" </pre>Install system packages: <pre>sudo apt update </pre><pre>sudo apt install -y python3-venv python3-opencv bluez v4l-utils </pre><pre>sudo systemctl status bluetooth </pre>Create a Python virtual environment inside the repo: <pre>python3 -m venv venv </pre><pre>source venv/bin/activate </pre><pre>pip install -r requirements-board.txt </pre>Find your USB camera: <pre>v4l2-ctl --list-devices </pre>In the examples below I use <code class="code">/dev/video0 </code>. Replace it with your actual camera device if needed. <h3>Download the trained model</h3>Run on: VENTUNO Q The fastest path is to use the trained ONNX model from the release: <pre>cd "$REPO" </pre><pre>source venv/bin/activate </pre><pre>mkdir -p models </pre><pre>curl -L -o models/speed_cnn.onnx \ </pre><pre>"https://github.com/munoz0raul/treadmill/releases/download/model-v1/speed_cnn.onnx" </pre>The model is about 44 MB. <h3>Run live AI speed estimation</h3>Run on: VENTUNO Q Start the live inference app: <pre>cd "$REPO" </pre><pre>source venv/bin/activate </pre><pre>python3 2-live-inference/live_speed.py \ </pre><pre>--camera /dev/video0 \ </pre><pre>--port 8090 \ </pre><pre>--cnn-model models/speed_cnn.onnx </pre>Open a browser at: <blockquote>http://&lt;board-ip&gt;:8090 </blockquote>You should see the live camera feed and a large speed readout. The estimator samples a rolling one-second video clip, runs the model with ONNX Runtime, applies a small motion gate for stopped-belt behavior, and smooths the result with a short rolling average. Demo video: https://youtu.be/9cKK7bAIcvc <h3>Run the Bluetooth treadmill app</h3>Run on: VENTUNO Q Stop the live inference app and start the Bluetooth FTMS app: <pre>cd "$REPO" </pre><pre>source venv/bin/activate </pre><pre>python3 3-bluetooth-ftms/game_server.py \ </pre><pre>--camera /dev/video0 \ </pre><pre>--port 8090 \ </pre><pre>--cnn-model models/speed_cnn.onnx </pre>Open: <blockquote>http://&lt;board-ip&gt;:8090 </blockquote>Press Start broadcasting . The board advertises itself as: <blockquote>AI Treadmill </blockquote>Now open your fitness game or BLE app and pair with AI Treadmill . In Zwift running mode, look under: <blockquote>Run &gt; Run Speed </blockquote>Important: many BLE fitness devices do not appear in the normal operating system Bluetooth settings. Pair from inside the fitness app, or use a BLE scanner such as nRF Connect or LightBlue. Bluetooth FTMS demo video: <a href="https://youtu.be/-HMXbTs5wZ4">https://youtu.be/-HMXbTs5wZ4 </a><h3>Verify BLE with the FTMS probe</h3>If a game does not pair, use the diagnostic probe from another BLE-capable computer or from the VENTUNO Q. <pre>cd "$REPO" </pre><pre>source venv/bin/activate </pre><pre>pip install -r requirements-debug.txt </pre><pre>python3 3-bluetooth-ftms/ftms_probe.py --name "AI Treadmill" </pre>A healthy run should find the device, connect, discover the FTMS service, subscribe to treadmill data, and show speed values similar to: <pre>[1] Scanning for 'AI Treadmill'... </pre><pre>PASS found device </pre><pre>[2] Connecting... </pre><pre>PASS connected </pre><pre>[3] Discovering GATT services... </pre><pre>PASS FTMS service 0x1826 present </pre><pre>[5] Subscribing to Treadmill Data 0x2ACD... </pre><pre>PASS subscribed </pre><pre>[7] Watching speed... </pre><pre>DATA 00002c01 speed=3.00 km/h </pre>The important characteristic is <code class="code">0x2ACD </code>Treadmill Data. Instantaneous speed is encoded in 0.01 km/h units. The app also exposes <code class="code">0x2AD4 </code>Supported Speed Range, because some games silently reject FTMS treadmills that do not provide it. <h3>Optional: train the model from scratch</h3>The first big lesson of this project was simple: good AI needs good data. The treadmill could not tell me its speed, so I built a data acquisition web app. I set the speed on the treadmill, tapped the matching speed button in the browser, and walked. Each click was saved with a timestamp next to the video. The public dataset contains 10 recording sessions and a prebuilt training cache. Run on: HOST machine <pre>git clone https://github.com/munoz0raul/treadmill.git treadmill </pre><pre>cd treadmill </pre><pre>export REPO="$PWD" </pre><pre>python3 -m venv venv </pre><pre>source venv/bin/activate </pre><pre>pip install -r requirements-train.txt </pre><pre>pip install gdown </pre><pre>gdown 1V_-AhkDP4gH7HobBH-CnSRrkIpcf5Evp -O dataset_repro.tar.gz </pre><pre>tar -xzf dataset_repro.tar.gz </pre>Rebuild the cache: <pre>python3 1-data-acquisition/preprocess.py \ </pre><pre>--sessions dataset_repro/session_* \ </pre><pre>--out dataset_repro/cache_side.npz </pre>Expected cache: <pre>X shape: (8679, 3, 8, 112, 112) </pre><pre>y range: 0.0 -&gt; 10.0 km/h </pre>Train and export: <pre>mkdir -p models </pre><pre>python3 1-data-acquisition/train.py \ </pre><pre>--cache dataset_repro/cache_side.npz \ </pre><pre>--out models/speed_cnn.pt \ </pre><pre>--warmup 5 \ </pre><pre>--finetune 25 </pre><pre>python3 1-data-acquisition/export_onnx.py \ </pre><pre>--ckpt models/speed_cnn.pt \ </pre><pre>--out models/speed_cnn.onnx </pre>I first tried to build a small video model myself, but my dataset was not large enough for that to work well. I looked for a model that already understood motion and used torchvision's <code class="code">mc3_18 </code>backbone pretrained on Kinetics-400 as the starting point. The classifier was replaced with a regression head that outputs treadmill speed in km/h. <h3>Optional: record your own dataset</h3>If your treadmill, camera angle, lighting, or walking style is very different from mine, you may get better results by collecting your own data. Run on: VENTUNO Q <pre>cd "$REPO" </pre><pre>source venv/bin/activate </pre><pre>python3 1-data-acquisition/acquire.py \ </pre><pre>--camera /dev/video0 \ </pre><pre>--port 8080 \ </pre><pre>--dataset dataset </pre>Open: <blockquote>http://&lt;board-ip&gt;:8080 </blockquote>Recording protocol: <pre>1. Set the treadmill speed. </pre><pre>2. Tap the matching speed button in the web UI. </pre><pre>3. Start recording. </pre><pre>4. Hold each speed for about 30 seconds. </pre><pre>5. Step through speeds from 0 to 8 km/h in 0.5 km/h increments. </pre><pre>6. Include at least one stopped or empty-belt recording. </pre><pre>7. Repeat across different days, people, clothing, and lighting conditions if possible. </pre>Then build a cache and train as shown above. <figure><img src="https://projects.arduinocontent.cc/0dab3374-1f26-489b-912c-038559b4038e.png"/><figcaption></figcaption></figure><h3>Optional: move the model from CPU to NPU</h3>The default path uses ONNX Runtime on the CPU. That is the easiest path to reproduce and debug. After the full camera to AI to Bluetooth loop was working, I also moved the same model to the VENTUNO Q's Hexagon HTP NPU using Qualcomm AI Runtime. The path that worked was: <pre>speed_cnn.onnx </pre><pre>-&gt; qairt-converter </pre><pre>-&gt; qairt-quantizer, A16W8 </pre><pre>-&gt; qnn-context-binary-generator for HTP V75 </pre><pre>-&gt; speed_cnn_a16w8_htpv75.bin </pre>The QAIRT SDK and runtime libraries are not bundled in the repository. They are Qualcomm-distributed SDK/runtime artifacts, and the compiled HTP context binary is also a generated file. The repository includes the full NPU build recipe so you can regenerate the artifacts with your own QAIRT install. Once the NPU runtime is staged under <code class="code">models/npu/ </code>, run: <pre>cd "$REPO" </pre><pre>source venv/bin/activate </pre><pre>python3 3-bluetooth-ftms/game_server.py \ </pre><pre>--camera /dev/video0 \ </pre><pre>--port 8090 \ </pre><pre>--provider qnn \ </pre><pre>--npu-runtime models/npu \ </pre><pre>--npu-strict </pre>Use <code class="code">--npu-strict </code>when recording a demo or benchmark. It disables automatic CPU fallback, so if the NPU runtime is not available, the app fails loudly instead of silently switching back to CPU. On 436 identical replayed clips, the quantized NPU model stayed within 0.054 km/h MAE of the float CPU output, while median inference latency dropped from 624 ms to 26.6 ms , about 23.5x faster . <figure><img src="https://projects.arduinocontent.cc/fe8c7994-7ca9-4f82-97c0-ac8368e51ec4.png"/><figcaption></figcaption></figure><h3>Troubleshooting</h3><h3>The camera does not open</h3>Check the camera device: <pre>v4l2-ctl --list-devices </pre>Then pass the right device: <pre>python3 2-live-inference/live_speed.py \ </pre><pre>--camera /dev/video2 \ </pre><pre>--cnn-model models/speed_cnn.onnx </pre><h3>The page opens but speed stays blank</h3>Check that the model exists: <pre>ls -lh models/speed_cnn.onnx </pre>If ONNX Runtime is missing, reinstall the board requirements: <pre>pip install -r requirements-board.txt </pre><h3>The game cannot find the treadmill</h3>- Start broadcasting from the web UI. - Pair from inside the game, not from the operating system Bluetooth settings. - Verify that BlueZ is running. - Run the FTMS probe. <h3>The speed reads nonzero when the treadmill is stopped</h3>The motion gate forces very-low-motion clips to 0.0 km/h, but camera noise, lighting flicker, or moving background objects can still affect the signal. Improve lighting, keep the camera fixed, and make sure the crop mostly contains the treadmill and walker. <h3>Safety notes</h3>This project does not control the treadmill. It only observes the belt and broadcasts an estimated speed. <ol><li>Do not adjust code, cables, or the camera while walking or running. </li><li>Keep cables away from the belt. </li><li>Use a stable camera mount. </li><li>Do not rely on this project for medical, safety, or emergency use. </li><li>Stop the treadmill if the game or web UI distracts you. </li></ol><h3>Source code and full documentation</h3>Full source, dataset instructions, NPU build guide, benchmark data, and article assets are in the GitHub repository: https://github.com/munoz0raul/treadmill <h3>AI collaboration note</h3>I have always loved having ideas. The difference now is that AI tools help me turn those ideas into working prototypes much faster. I do not want to pretend that I typed every line of this project alone. I used AI as a collaborator for code, debugging, documentation, and iteration. At the end, I personally reproduced the scripts and commands to validate the process. <blockquote>Note: The VENTUNO Q shown in this project is a pre-production unit. The final production hardware may differ in appearance and configuration. </blockquote> </body></html>

Give Your Treadmill Eyes with Edge AI

Turn any treadmill into a Bluetooth smart treadmill using a camera and Edge AI running locally on the Arduino VENTUNO Q.