Bird Feeder Species Logger on Raspberry Pi

Backyard birdwatching generates wonderful data that nobody records: which species visit, when, and how often. Commercial smart feeders exist but they are closed, expensive, and useless as student projects. This build does it the open way: a wooden feeder fitted with a Raspberry Pi 4 and Pi Camera module, watching the seed tray. A PIR sensor plus frame-differencing wakes the capture pipeline only when something moves; the captured frame is cropped to the bird and passed through a MobileNetV2 convolutional network fine-tuned on a 20-species labeled bird dataset, and the top prediction with its confidence is written to a local SQLite log with the photo. The dashboard — the software half of this hybrid project — renders the detection feed, today's visit counts, an hourly activity chart, a species breakdown and a review queue for low-confidence frames. The honest-ML discipline matters here: the notebook trains on a documented dataset split, evaluates top-1 accuracy on a held-out validation set, and the report discusses the real failure modes — squirrels, motion blur, harsh shadows and species outside the training set, which are flagged for review rather than mislabeled.

1. The PIR sensor and frame-differencing detect motion at the tray and trigger a 1080p capture from the Pi Camera.
2. The frame is cropped around the motion region and resized to the classifier's 224x224 input.
3. The fine-tuned MobileNetV2 model runs on-device and returns the top species prediction with a confidence score.
4. Predictions above the confidence threshold are logged with timestamp and photo; below-threshold frames go to the review queue.
5. The dashboard reads the SQLite log and renders the detection feed, hourly chart and species breakdown.
6. At week end, the log exports to CSV; the buyer can review flagged frames and correct labels.
7. The training notebook lets the buyer fine-tune on additional species and re-run the validation evaluation.

Tech stack:

• Raspberry Pi 4 (4 GB) + Pi Camera Module v2
• PIR motion sensor, wooden feeder with camera bracket
• Python, TensorFlow Lite (MobileNetV2 fine-tuned, 20 species)
• SQLite detection log
• Single-file HTML/CSS/JS dashboard (feed, charts, review queue)
• Training notebook (dataset prep, fine-tuning, evaluation)

• [Motion-triggered capture] PIR sensor plus frame-differencing wakes the camera only on movement, saving storage and compute.
• [On-device species classifier] MobileNetV2 fine-tuned on a 20-species labeled bird dataset; inference runs on the Pi 4 in about 1.4 s per frame — no cloud needed.
• [Detection feed dashboard] Latest detections with species, confidence, timestamp and photo; today's visit count and hourly activity chart.
• [Species breakdown analytics] Visits per species over the week, top visitor, rare-sighting highlights and peak-hour analysis.
• [Manual review queue] Frames below the confidence threshold are flagged for human review instead of being silently mislabeled.
• [eBird-compatible export] Weekly logs exportable as CSV with species, counts and timestamps for citizen-science use.
• [Retraining notebook] Documented fine-tuning pipeline so the buyer can add local species and re-evaluate — accuracy is measured, not pre-claimed.

• Bird feeder hardware with Pi Camera bracket, PIR sensor and Raspberry Pi 4 assembly
• Complete capture + classification software (motion trigger, inference, logging)
• Fine-tuned MobileNetV2 model (TFLite) + training notebook (dataset prep, fine-tuning, evaluation)
• Detection dashboard web app (feed, charts, species breakdown, review queue, export)
• Wiring and mounting guide (feeder placement, camera angle, weather protection notes)
• Project report PDF (background, dataset analysis, methodology, buyer-run evaluation, error analysis)
• PPT presentation for final review
• Viva Q&A preparation document (CNNs, transfer learning, TFLite, evaluation metrics, dataset bias)
• Setup guide (OS image, dependencies, first capture test)

• Classification accuracy is a design target (~85% top-1), not a guarantee — the notebook measures it on the buyer's validation split and the report documents the real number, including confusions between similar species.
• Night and low-light frames are unreliable (no IR illumination in the base build); the pipeline skips dark frames and logs the skip.
• Squirrels, falling leaves and wind-shaken branches trigger captures — about one in ten triggers is a false positive, filtered by a motion mask the guide documents.
• Species outside the 20-class training set cannot be recognized; they land in the review queue, which is the honest behavior, not a failure.
• Rain and feeder placement affect results; the mounting guide covers siting, and weatherproofing the Pi is the buyer's responsibility.
• Continuous inference shortens SD-card life; the trigger-gated design exists precisely to limit writes.