Camera-based
driver behaviour analysis
Challenge
The goal of the project was to create next-generation driver behaviour analysis platform for motor insurance purposes. Unlike all other existing platforms, the solution should utilise truly contextual data, coming from a camera. Additionally, the solution should be cheap to introduce to B2C customers, thus it cannot require additional hardware.
Tool used
Tensorflow, Keras, YOLO, CNN
Solution
We have created a module that uses images from smartphone camera to analyse dangerous road situations, such as tailgating, hopping between lanes of traffic or speeding in a proximity of pedestrian crossing. We have built a deep learning model, using CNN model and YOLO detector. We have used transfer learning from a model trained on COCO and then trained it on a custom-built dataset (ca 50 000 images captured from a dash camera) on our specific use-case.
See also
