Possum Tracker
Real-time CNN possum detection for backyard wildlife monitoring
A computer vision pipeline that detects possums in night camera footage using motion analysis and CNN classification via transfer learning. Built to prevent wildlife-pet conflicts by triggering smart home automation when a possum is detected.
How It Works
01
Camera Feed
02
Motion Detection
03
ROI Extraction
04
CNN Classification
08
Dashboard
07
Data Storage
06
Alert
05
Sliding Window
01
Camera Feed
02
Motion Detection
03
ROI Extraction
04
CNN Classification
05
Sliding Window
06
Alert
07
Data Storage
08
Dashboard
Model Performance
99.7%
Test Accuracy
100%
Precision
99.6%
Recall
3,061
Test Samples
Skills Demonstrated
Computer Vision
Motion detection via background subtraction to extract regions of interest from RTSP camera feeds, handling noisy night footage with insects, shadows, and rain.
Deep Learning
Transfer learning with ResNet18 for binary possum classification. Custom classification head trained on motion-extracted ROIs with padding-based resizing to 224x224.
Data Collection & Preprocessing
Automated ROI generation from night camera recordings with session-based train/test splitting to prevent temporal data leakage. Manual review, sorting, and labeling.
Full-Stack Web Development
This portfolio site built with a modern React framework (with support from Claude AI), server-side rendering, component-driven architecture, and a MySQL backend.
Real-Time Systems
Live camera feed processing with frame skipping, batch-based training, and per-ROI inference. Temporal sliding window (3/5 frames) to stabilize predictions.
Infrastructure
Containerized development environment, CI/CD pipelines, and database-backed analytics for logging possum visit timestamps and detection events.