POITIVE is a parking space availability computer vision project. Trained Yolov8's object detection CV model on a custom dataset containing multiple images of neighborhood curbs and relative objects (orange box). Actively look for open curbs and measure via a detected relative object. Terminal interface that runs the model given a specified time interval to detect for open parking space.
- Reduced input size (imgsz) to a multiple of 32 due to backbone filter stride of yolov8 architecture. Made sure dataset images follow same dimensional scheme.
- Training terminated at 100 epochs
- Model performance:
- mAP50: 0.74112
- val/cls_loss: 1.432
The project provides a terminal interface that runs the model at specified time intervals to detect open parking spaces. Notifications sent through text. Create a .env file containing sender password/email.
python3 interface.py
-Camera parallel to the curb case, implement an OBB yolov8 model instead to obtain 4 corners of bounding box. Utilize dist. formula for curb space calculation.
-Varying perspective case, where the curb seems like its minimizing to a certain point, make use of 2 relative objects of the same height for calibration. Assume linear perspective and apply a linear transformation to length calculation.
-Expand dataset.
-Retrain model across varying initializations/epochs.
-Expand use to cases from neighborhood curbs to parking structures/lots.
-Create API endpoints and establish a cleaner front facing user interface.