Parking occupancy from a camera — no per-space sensors required

DAPS instance-segments occupied vs. open parking spaces straight from existing lot cameras, replacing per-space hardware.

Role

Solo — end-to-end: data, training, evaluation

Goal

Detect open spaces without installing per-space sensors

Key outcome

0.86 avg IoU · 0.90 precision/recall

Stack

Python · PyTorch · Mask R-CNN (ResNet50-FPN)

0.86

average IoU on occupied vs open space segmentation

0.90

precision and recall — reliable enough for live guidance

12,416

lot-camera images used for training and evaluation

$ cat problem.md

Hardware sensors don't scale; cameras are already there

Per-space occupancy sensors cost money for every single spot. Lots already have cameras — the problem is turning oblique, overlapping views into per-space occupancy, which needs instance segmentation, not just detection.

$ ./approach.sh

01Adapted Mask R-CNN (ResNet50-FPN) to instance-segment occupied vs open spaces from lot cameras.

02Trained and evaluated on 12,416 images of real lots.

03Reached 0.86 avg IoU and 0.90 precision/recall — with zero added hardware per space.

$ cat results.json

0.86 average IoU, 0.90 precision/recall on occupied-vs-open segmentation
Works from existing lot cameras — no per-space sensor install
Solo project — full pipeline from data to evaluation

Parking lot with per-space occupancy — colored instance masks, space-empty / space-occupied labels

Fig 1 — per-space occupancy from a single camera view

Wide parking-lot view with instance masks and occupancy confidence scores

Fig 2 — dense lot, occupied vs open with confidence scores

$ grep -r skills

instance segmentationMask R-CNNResNet50-FPNPyTorchdataset curationmodel evaluation