Impulse Geophysics

Video Redaction

​

Redaction by masking Personally Identifiable Information (PII)

 

There are two primary redaction steps - the localisation of PII object(s) and their obfuscation. In the method used by Impulse Geophysics the localisation is also  in two steps – locate people and then locate faces. Using this method allows a person to be obfuscated from all angles back, side and front with almost 100% redaction – beyond the Google Streetview standard.

​

Detection and Recognition are both achieved by using deep neural network based models.

​

 

​

Redaction of Number Plates

​

The detection of number plates is a common goal in computer vision. There are many ANPR cameras in operation; in police vehicles and statically mounted cameras used for speed monitoring, tax enforcement and parking enforcement.

​

From a machine learning perspective, detection of license plates is reasonably straightforward – the by-design uniformity of a license plate makes for a consistent pattern that a machine learning algorithm can train on. Specifically, in the UK, license plates are yellow or white rectangles with one or two rows of letters or numbers.

​

To detect number plates in real world street scene imagery is, however, a more complex problem. Number plates on a vehicle directly in front of the camera may be easy to detect, but number plates on parked cars may be partially occluded or present at a significant angle (for example, a car waiting to turn out of a side road), which will confuse the detection algorithm.

​

After considerable experimentation it was found that the best approach to detecting number plates was two parts: first examine each image for white or yellow regions, and then attempt to detect characters within the regions using optical character recognition.