How this demo works
This demo estimates vehicle counts from ambient audio alone — no camera or radar needed for detection. Pick a sample clip from the sidebar and run the model.
The video is shown only so you can visually check the result against what's happening in the clip — the model itself only ever sees the audio track, never the video frames.
Results are labeled Model Prediction — the model's estimated count, split left→right / right→left, with a calibrated confidence range in parentheses where available. Below the video you'll also see a separate Ground Truth block with the manually annotated true count for that clip, so you can compare the prediction against it directly.
The mel spectrogram is the same time-frequency view the model sees — useful for a sanity check against audible events (engine passes, horn honks) in the clip.
1 — Load clip & select model
2 — Model Prediction Results