INTERNSHIP - Face blurring on train CCTV images

Modern trains are usually equipped with CCTV cameras that continuously monitor the status of the infrastructure and the behavior of the passengers. When damage or suspected behavior is detected, the video feed can be reviewed by the train staff or by law enforcement to investigate the issue.

In recent years, demand has grown for ML/AI solutions to automatically detect possible anomalies. The advantages are clear: defects and problematic behavior can be detected in real-time and the train staff can intervene before further damage happens. To train these tools, however, a dataset with thousands of images is required, from as many different angles/trains and for as many different situations as possible. Collecting video data from unsuspecting passengers is not allowed under GDPR, and obtaining approval from individual passengers is unfeasible. Therefore, Televic GSP would like to automatically blur all passenger faces in the training data, such that no passengers are identifiable, but the unblurred images should still be available for training purposes.

The student should

* investigate different blurring methods and compare them with respect to accuracy, speed, ease-of-use, etc.

* build a proof-of-concept tool that automatically extracts blurred images from a video stream, lets the user label the blurred images, and transfers the labels to the unblurred images for training.

If you are interested in this topic, please also register this on the Televic website at: https://www.televic.com/en/careers/internships-and-students so we can confirm the topic is still available.