Clean masters

The goal of this work is to define a machine learning pipeline capable of detecting dirt in vehicles in order to advise customers of the possibility of cleaning their car in the establishment and thus increase the use of the cleaning services of the different GALP gas stations.

We have decided on a multimodal approach that encompasses object detection tasks 📸, semantic segmentation 🧩 for the separation of the different parts of a car (rear-view mirrors, wheels, hood, etc.) and classical computer vision methods🧼 that are able to independently evaluate the parts detected in the segmentation process and detect the dirt points with a high level of granularity.

🚙 Car Detection 📸

File: car_detection.ipynb

The camera is always recording, but the Dirt Detection Pipeline is not to be activated until a car is detected, therefore the YOLO Model is used which can detect cars frame by frame with a high speed and serve as a "switch" of the whole system when a car is detected.

🚙 Car parts detection 🧩

File: car_parts_seg.ipynb

In this file he has used a semantic segmentation approach to detect the different parts of a car to facilitate the detection of the different dirt parts of a car.

🚙 Car parts dirty detection 🧼

File: Dirty_detection_image_processing.ipynb

In this file, the detection of dirt on the different parts of the car that have been previously detected in car_parts_seg is performed.

🚙 CNN model 🧠

Finally a CNN model was developed and trained to detect if a car is clean or not, with a 92% of validation accuracy and 88% test accuracy.

🚨 As part of the first part we have defined the detection pipeline and the image processing by evaluating each of the three components of the pipeline separately. After having obtained the first results, in the next phase, we will proceed to the integration of all elements of the pipeline.