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.
This system was inspired by consistent nocturnal visits from a backyard brushtail possum captured on a home security camera. The goal was not only to detect presence in real time, but to extract measurable behavioural features and analyse patterns across visits.
The result is a combined detection and behavioural analytics pipeline operating on real-world, noisy night-time data.
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
Test Accuracy
Precision
Recall
Test Samples
Motion detection via background subtraction to extract regions of interest from RTSP camera feeds, handling noisy night footage with insects, shadows, and rain.
Transfer learning with ResNet18 for binary possum classification. Custom classification head trained on motion-extracted ROIs with padding-based resizing to 224x224.
Automated ROI generation from night camera recordings with session-based train/test splitting to prevent temporal data leakage. Manual review, sorting, and labeling.
This portfolio site built with a modern React framework (with support from Claude AI), server-side rendering, component-driven architecture, and a MySQL backend.
Live camera feed processing with frame skipping, batch-based training, and per-ROI inference. Temporal sliding window (3/5 frames) to stabilize predictions.
Containerized development environment, CI/CD pipelines, and database-backed analytics for logging possum visit timestamps and detection events.
Visit-level behavioural metrics computed and aggregated via database queries and backend logic. Statistical correlation analysis and interactive visualisations delivered through API-driven dashboard components.
Large-scale biodiversity occurrence data processing and spatial aggregation using hexagonal grids. Automated ingestion of observation records from the Atlas of Living Australia API, incremental database updates, and interactive visualization of species observation density across Australia.
