Recently, a few companies in the body-worn camera space have been marketing what they consider to be fully automated video redaction capabilities. While we applaud efforts to work toward a better video redaction solution for law enforcement agencies, it’s important to shed some light on why some of their claims are untrue. In addition, while some believe that full automation of video redaction would help agencies handle Freedom of Information Act (FOIA) requests more efficiently, we feel it’s necessary to discuss the inherent pitfalls with that approach.
What are the challenges of automated video redaction?
1. Redaction is contextual.
Understanding the context of a scene requires human thinking. A computer is not intelligent enough to know which faces or objects in a video need to be redacted. Perhaps you only want to blur out the face of witnesses and juveniles in a scene, but want to show the face of a suspect. Or, maybe certain documents or IDs are shown in a video, but you only want to redact the items that contain personally identifiable information. What may appear as a simple task for a human to comprehend is a very complex problem for a computer to understand.
In a debate posted on Urban.org, Spokane Police Chief, Frank Straub, said, “I’m not convinced that redaction can be completely automated. There are things such as in-car computer screen images that have to be recognized and redacted, uninvolved persons, potentially background that could identify a victim, etc.”
Nancy La Vigne, Director of the Justice Policy Center at the Urban Institute, agreed, “I share Chief Straub’s skepticism about automated redaction – even if it were possible, it would be imperfect. Plus, any redaction also results in the loss of critical context with which to understand the officer-citizen interaction.”
2. Computer vision has limitations.
Our company, MotionDSP, has been developing image processing and computer vision software for more than 10 years, working with some of the toughest video from the military. We understand the complexity, challenges, and promise of computer vision-assisted video technology, but also understand its limitations.
Computers are not very intelligent when it comes to recognizing things in video. Tasks like recognizing faces are much easier when the variables in your video are consistent. For example, if you point a security camera at a metal detector that people walk through, you can guarantee faces will be looking forward and in roughly the same position each time. Significant challenges arise when you introduce multiple moving variables as found in body camera footage. Body cameras are attached to people that walk, run, jump, crouch, and sit. People or objects being filmed can also be moving in the scene. This makes it extremely difficult to predict how a face or object will look in a video from a body camera versus a stationary camera.
Auto-detecting a face can be relatively simple if you have a perfect scenario as described above. You look for an oval-like shape that has skin tones, two eyes, a nose, and a mouth. However, a number of problems can arise with this method:
- What if the person turns to the side? Their face is no longer an oval, and you can no longer see both their eyes.
- What if the scene is at night? The skin color is now going to be vastly different. If the camera switches to infrared mode, there won’t be any skin tones. What happens if it is so dark that there’s not enough contrast between a person’s head and the background?
- What if the police officer is walking or running? In this case, the face you are trying to redact will be moving up and down, possibly at a rapid pace. How do you track that reliably?
- What if all three of the above happen simultaneously? Can we really expect automated detection and redaction to work?
3. Zero margin for error.
If a tracking algorithm fails for even a single frame, a face or personal information can be revealed, completely defeating the purpose of redaction. We’ve seen this happen in marketing videos of redaction products of more than one company that claim their solutions are fully automated. The blur is tracking along fine, then a person makes a sudden movement and the blur no longer covers their face. If this happens in prepared marketing examples using controlled environments, just imagine the issues you’d run into with more complicated videos.
4. Accept that humans need to be in the loop.
Public safety is a complicated profession and becomes even more challenging when you consider all of the complexities of privacy concerns and handling sensitive information. While technology likely won’t ever be a substitute for human intuition, it can simplify mundane, time-consuming processes and allow law enforcement professionals to spend more time on things that require their insights and expertise.
If you’re curious about our approach to redaction, be sure to check our product, Ikena Spotlight.