- Using a Raspberry Pi + webcam, record the speed of the vehicles driving in my street.
- Label the training set using my own tool (Mechanical French)
- Then, apply an image classifier (Tensorflow convnet) to generate stats per vehicle category.
- Finally, implement a Dashboard for the processed data & Look for correlation with other data (weather, time slot, events)
- For speed calculation, on the shelf library speedcam, using opencv mean_shift method.
- For vehicles classification : standard keras wrapper on tensorflow with image augmentation, applying standard CNN layers layout.
- 30000 vehicle's speeds recorded, 18000 manually classified.
- Validation accuracy on a binary classifier (car vs bike) : 97%