This Jupyter Notebook trains and evaluates a binary image classifier to determine whether a parking spot is occupied or free using video input. The pipeline uses TensorFlow with MobileNetV2 via transfer learning and fine-tuning. It also includes video frame extraction, contour-based segmentation, and frame-by-frame classification visualization.
- Preprocessing of labeled image data (empty vs not_empty)
- Image augmentation for improved generalization
- Model training using MobileNetV2 with early stopping and learning rate scheduling
- Fine-tuning of the top layers of the pre-trained model
- Evaluation with accuracy, loss, and a classification report
- Frame extraction from video using a binary mask of parking spots
- Cropping, resizing, and classification of each parking spot in selected frames
- Visualization and overlay of predictions on original video frames
- Python 3.8+
- TensorFlow 2.19+
- OpenCV
- NumPy
- Matplotlib
- Seaborn
- scikit-learn
Structure your dataset as:
clf-data/
empty/
not_empty/
Run the notebook cells up to training and fine-tuning the model.
Accuracy, loss, confusion matrix, and a classification report are generated to evaluate the model.
Provide a parking lot video and a corresponding binary mask. The notebook extracts frames, crops individual spots, classifies each, and overlays predictions.
A video with drawn bounding boxes for each spot labeled as either free (green) or taken (red) is generated.
CUDA and cuDNN setup is optional but recommended for faster training. This notebook is designed for educational and prototyping purposes. Real-world deployment may require additional calibration, robust detection pipelines, and environmental variability handling.