How AI Detected This Animal
The same pipeline runs for uploads, camera-trap batches and the live camera feed.
- Camera / Image
- Preprocessing
- AI Model
- Object Detection
- Species ID
- Counting
- Analysis
- Speech
1 · Image
A photo arrives from a camera-trap file, an upload or a frame grabbed from the live camera stream.
2 · Preprocessing
The image is resized (long side ≤ 1024 px) and compressed in the browser so inference is fast and bandwidth-light.
3 · Object Detection
The detection model scans the image and proposes regions that contain animals — not one label for the whole picture.
4 · Bounding Box Generation
Each proposal becomes a box in normalized coordinates (0–1), so it scales perfectly at any display size.
5 · Species Classification
Every box is assigned a class label from the configured species list, or 'unknown' when the model is unsure.
6 · Confidence Score
Detections below your threshold are discarded, and overlapping boxes are merged with non-maximum suppression (IoU).
7 · Object Counting
Surviving detections are grouped per species to give per-species counts and a total animal count.
8 · Text-to-Speech
Counts are converted into a grammatical English sentence and spoken with the Web Speech API.
Why object detection, not image classification?
An image classifier gives one label per photo — it cannot tell you that a frame holds two deer and a wild boar. Object detection localises each individual animal, which is what makes counting, density estimates and camera-trap studies possible.
Non-maximum suppression removes duplicate boxes for the same animal by discarding lower-confidence boxes whose intersection-over-union with a stronger box exceeds the IoU threshold.