Parking Lot Occupancy Detection using Computer Vision

Drivers circling for parking waste fuel and time, and lot operators have no real-time picture of where the free bays actually are. Sensor-per-bay hardware exists but is expensive to install and maintain. This project does it with the cameras already on the poles: a CNN classifier trained on PKLot (12,417 lot images with ~696,000 segmented parking spaces, across three lots and varied weather) labels each marked bay as occupied or empty from an overhead frame. Zone tallies feed guidance signage and a driver-facing availability view. The demo dashboard (the screenshots) shows the live lot with per-bay overlays, an occupancy donut, zone-wise bars, hourly trends and dwell-time stats. Built with TensorFlow/Keras and OpenCV, the pipeline — bay marking, classification, aggregation — is fully explainable in the viva.

1. The operator marks each parking bay as a polygon once per camera view; polygons are saved as JSON.
2. Each new frame is warped to the stored view (optional stabilization), and every bay polygon is cropped.
3. Each crop is resized to 128×128, normalized and passed through the binary CNN — occupied or empty with confidence.
4. Per-bay results are aggregated into zone tallies and lot-wide occupancy.
5. Zone tallies are formatted for guidance signage or an app feed.
6. The dashboard renders bay overlays, the occupancy donut, zone bars and the hourly trend chart.
7. Analytics accumulate dwell times and turnover statistics from the timestamped bay history.

Tech stack:

• Python 3, TensorFlow/Keras
• OpenCV (bay polygons, cropping, stabilization)
• PKLot dataset (12,417 images, UFPR)
• NumPy, Matplotlib (evaluation plots)
• ONNX Runtime (edge deployment)
• HTML/CSS/JS single-file dashboard (this demo)

Dataset & model details

• Dataset: PKLot (de Almeida et al., UFPR) — 12,417 parking-lot images with 695,851 segmented parking spaces across three lots (UFPR04, UFPR05, PUCPR), covering sunny, cloudy and rainy conditions.
• Task: Binary image classification per bay crop; input = 128×128×3 crop, output = occupied/empty probability.
• Model: Custom CNN (approximately 900k parameters): conv blocks 32→64→128, global average pooling, Dense(128, ReLU), Dropout(0.4), sigmoid output.
• Metrics: Bay accuracy 99.1%, cross-lot accuracy 96.8% (train on two lots, test on the third), rain/night drop under 2.5 percentage points — design targets for the built-to-order training run. The report documents the actually achieved figures.

• [Per-bay occupancy CNN] Binary classifier (occupied/empty) trained on ~696,000 PKLot segmented spaces, applied to each marked bay crop.
• [One-time bay marking] Operator marks bay polygons once per camera view; the tool stores them as reusable JSON — no per-frame manual work.
• [Live lot dashboard] Overhead view with per-bay occupied/free overlays, occupancy donut and free-bay count in a single-file web UI.
• [Zone analytics] Zone-wise occupancy bars, hourly occupancy trends, peak detection, average dwell time and turnover-per-bay statistics.
• [Guidance output] Zone tallies formatted for signage or a driver app feed ("Zone C: 41% — 38 bays free").
• [Weather-robust training] PKLot's sunny, cloudy and rainy subsets are all in training; the report measures the accuracy drop per condition.
• [Edge-ready export] Classifier exports to ONNX so a pole-side device can score a full lot locally.

• Trained bay-classifier model and the complete training notebook
• Bay-marking tool + occupancy aggregation Python pipeline
• Single-file HTML lot dashboard wired to pipeline outputs
• Per-condition accuracy analysis and trend charts for the report
• Project report PDF (background, dataset, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (binary classification, PKLot, cross-lot generalization, deployment)

• Bay polygons must be re-marked if a camera is moved or re-aimed; the marking tool makes this a minutes-long job.
• Heavy snow or flooding occluding bay markings degrades accuracy — documented as an environmental limit.
• 99.1% bay accuracy is a design target — the report documents the actual achieved figure, including per-condition breakdowns.
• Night operation needs adequate lot lighting or IR illumination.
• As shipped, guidance is informational; barrier/payment integration is future scope.