BirdCamAI

Developed a real-time bird detection and identification system capable of accurately detecting and classifying approximately 900 bird species with an average latency of ~400ms. Upon detection, the system automatically sends an email alert.

System Architecture

Picamera2 → Capture frame

↓

NCNN INT8 YOLOv11n (~300 ms)

↓

Postprocess (NMS, confidence threshold)

↓

Bird detected?

├─ YES → Lookup species → Save image → Email notification

└─ NO → Discard frame

↓

Sleep / next capture

Model Architecture & Training

Base Model: YOLOv11n (nano variant) configured for single-class bird detection with 320×320 training resolution.

Training Setup:

- Batch size: 32

- Dataset: Google open image dataset V7 (~27,000+ annotated bird bounding boxes across diverse backgrounds like water, grass, rocks, sky and more)

Performance Metrics

Edge Optimization Pipeline

Stage 1: Model Export

After training, the YOLOv11n model was exported from the training framework and converted into NCNN format (.param + .bin), which is designed for efficient inference on ARM CPUs.

Stage 2: INT8 Quantization in NCNN

NCNN’s quantization tooling was used to convert the model to INT8 weights and activations:

Compression: FP32 → INT8 (roughly 4× smaller model).
Speedup: around 1.3–1.4× faster on the Pi Zero 2W CPU compared to FP32 NCNN.
Accuracy impact: <5% relative drop (mAP@0.5 from ≈0.79 to ≈0.75).

Stage 3: Runtime Optimization on Pi Zero 2W

On-device, inference is performed using the NCNN C++/C API (or its Python binding, depending on setup):

Picamera2 captures frames and resizes them (e.g., to 224×224 or another tuned size).
The NCNN pipeline performs preprocessing, INT8 inference and postprocessing (decoding + NMS).
Only a single thread is used to match the Pi Zero’s limited CPU resources.

Result: roughly 450ms end‑to‑end latency per frame (capture → NCNN inference → postprocess).

Evaluation & Insights

Precision-Recall & Confidence Curves

The model maintains high precision (~0.80) across a wide confidence range, with recall gradually dropping as threshold increases. This suggests the model is well-calibrated and benefits from careful threshold tuning rather than retraining.

Confusion Matrix

True positives: 84% (normalized).

False negatives: 16% (birds missed at low confidence).

False positives: <1% on backgrounds (grass, water, rocks).

The model rarely "hallucinates" birds, which is critical for a bird feeder camera—fewer false alarms mean more reliable notifications.

Conclusion

This project demonstrates that sophisticated computer vision—reliable bird detection with 0.79 mAP—can run continuously on a $15 microcomputer without cloud infrastructure or GPU acceleration. The combination of model selection (YOLOv11n), aggressive quantization, and careful edge optimization bridges the gap between research-grade accuracy and production-grade efficiency on ultra-constrained hardware.

Page updated

Google Sites

Report abuse