Counting what off-the-shelf models can't see.

An FDOT-funded safety study needed reliable counts of bicycles, e-bikes, and e-scooters at fixed camera sites across the district. Manual counting doesn't scale past a handful of videos, and stock detection models kept folding the interesting classes into each other: cars read as bicycles, e-bikes read as motorcycles, seated scooters read as almost anything.

Two passes: catch everything, then trust nothing.

A motion pass (MOG2 background subtraction inside a hand-drawn region of interest per site) catches every moving object cheaply. A review pass filters that to real micromobility before anything reaches the dataset. Production counting runs YOLOv8 for detection and ByteTrack for tracking, so each rider is counted once, not once per frame.

5,000+ frames, labeled by hand.

The training set merges field footage with public datasets, plus negative classes (cars, motorcycles) so the model learns what a bike is not. Missed cases get mined back in: a frame-extraction pipeline reads burned-in video timestamps via OCR, pulls the exact frames where the detector failed, and routes them through Label Studio with pre-labels for fast annotation.

Nine classes, production grade.

Iterating on the detector with negative classes and mined hard cases pushed the toughest distinction, e-bike vs seated scooter, to ~0.90 mAP, eliminated false motorcycle detections, and was verified head-to-head on a held-out site before delivery.

Counts are the product. Safety is the story.

Every verified detection carries a hand-reviewed set of safety attributes before it counts:

Shipped like a product, not a notebook.

No model version ships on faith: a manual review pass screens detections, and each new detector is run head-to-head against the previous one on a held-out site before it replaces anything. What the research team receives is reproducible, not a one-off run:

Sharper eyes without new satellites.

Most freely available satellite imagery sits around 10 meters per pixel: enough to see a field, not enough to trace building footprints, assess crop health, or map flood lines. Higher-resolution acquisition is expensive or simply unavailable. Computational super-resolution upgrades the imagery that already exists.

Two networks, locked in an argument.

An SRGAN: a generator (16 residual blocks feeding 4× upsampling stages) learns to hallucinate plausible high-resolution detail, while a discriminator learns to call its bluff. Instead of pixel-only loss, a pretrained VGG19 network supplies perceptual loss, comparing feature maps rather than raw pixels, so the output gets real texture and sharp edges instead of the smooth blur that pixel losses produce.

5,000+ aligned pairs, GIS-ready in and out.

Training ran on 5,000+ spatially aligned high/low-resolution GeoTIFF pairs spanning varied terrain. A rasterio preprocessing pipeline stacked RGB bands, normalized intensities, and filtered corrupted files; low-resolution inputs were generated by controlled downsampling so every pair aligns perfectly. Geospatial metadata is preserved through the whole pipeline and reinstated on outputs, so enhanced imagery drops straight into GIS tools.

Numbers that hold up.

The project report logged 27.2 dB; continued training after the report pushed the final generator to 31 dB. Pixel-error heatmaps localize exactly where the generator struggles (dense urban edges, fine vegetation texture), turning failure analysis into a map instead of a guess. And the honest trade-off, stated plainly: GANs trade a little raw PSNR for perceptual realism. For imagery that humans interpret, that is the right trade.

Five 12-hour epochs on a single GPU.

Adversarial training is heavy: five epochs at 12+ hours each, roughly 60 GPU-hours on an RTX 3060. That forced disciplined batch sizing, checkpointed generators, and deliberate experiment planning rather than spray-and-pray runs. Current scope is RGB; extending to multispectral bands (NIR, SWIR) is the natural next step for vegetation and water analysis.