The Power of Machine Learning in Automated Video Redaction

The Power of Machine Learning in Automated Video Redaction

Many people are aware of AI or Artificial Intelligence and its meaning, especially in the way that it is often portrayed through movies. These movies are often exciting and captivate our imaginations. Machine learning, while similar to AI, is defined differently. A way to explain this in layman’s terms is that AI is the breadth of knowledge contained and used by a system, while machine learning is the algorithms or processes in which the system gains the knowledge and assimilates it for future use. In human terms, AI would be all the information and knowledge you already have, while machine learning would be likened to the steps you choose to acquire that knowledge, such as reading, observing, studying, or even making mistakes.

Understanding Machine Learning

A good definition of machine learning would be the concept that by feeding data to a machine it can learn over time to anticipate and make decisions on its own and act accordingly by using the information on hand from observations and real-world interactions. Using pattern recognition as a guide, computer software uses the data entered to learn to perform tasks independently. Many industries today rely on machine learning to mine their large amounts of data in order to work more efficiently and to acquire an edge over their competition.

Law enforcement and other government agencies including Homeland Security rely on machine learning software to better make accurate real-time decisions and keep the public safe. Police officer’s own safety may rely on the abilities of machine learning or data mining. The decision-making process that goes into their video redaction software used with their personal cameras is linked to internal systems which can give them immediate access to information gathered from multiple sources. The accurate information returned gives the officer prompt, secure delivery while giving them the intelligence they need so that they can make quick decisions while continuing to follow their policing policies and practices during which time they are also able to act on threats to the public in a reliable and safe manner.

Supervised Learning Vs. Unsupervised Learning

There are largely two main methods of machine learning, supervised learning and unsupervised learning. These are not the only methods, but the two most popular. Reinforced learning is also another method for machine learning and each technique provides a different approach to learning, gathering data, and generating informative results. It largely depends on the industry which approach is used, but today as many systems become even more cross-integrated, newer software designs apply multiple machine learning methodologies.

All machine learning methods use algorithms to determine solutions, gather information, and discover the results required to create an action or solve a problem. When the technique for gathering information is done with supervised learning the algorithms are trained using patterns and models that are labeled when input as data. Remember that algorithms are simply a series of steps or methods used to solve the problem. With supervised learning, the software obtains a specific set of input matched up to a distinct set of output. The way that the machine begins to learn to discover errors and initiate its own problem-solving skills is by comparing the original correct output with its present output. This is machine learning that will build upon classification, labels, regression, and prediction sequences to locate errors and predict future values. This type of machine learning is generally used in applications that can compare histories and predict future behaviors. A common use would be having a software application that can use intuition to anticipate fraud or other criminal activities in banking and credit card transactions. It can also be used to predict which customers may be likely to file insurance claims for damages on homes and autos.

Unsupervised learning is a good tool for applications involving marketing and gaining population samples for studies. Principal component analysis and cluster analysis are common methods used to acquire data sets using unsupervised machine learning. Unsupervised learning works by extrapolating data sets that are found to have similar properties. By studying the unlabeled data, the software application searches for commonalities and groups the data into clusters for further inquiry. It gathers the data into usable sets because they have not been previously labeled, classified, or categorized. An example would be finding and grouping sets of customer data based on shopping habits so that marketing material can be targeted to their needs and more effective in creating purchases; which in turn generates a higher profit margin.

Other types of machine learning are used for other purposes and more methods are constantly being developed. Two more but less commonly used learning methods are semi-supervised learning and reinforcement learning. Semi-supervised learning is often used in the same types of applications using historical data and predictions as supervised machine learning techniques. The difference is that it uses both labeled and unlabeled data points to learn to make classifications and future predictions. This is a good option for many businesses as unlabeled data points can be delivered in large quantities and generally for less money than their labeled counterparts. Reinforcement learning is different in that it allows the algorithm to learn through trial and error which operations to execute to achieve the greatest yield or gains. The maximum reward over time is the general objective. This time of machine learning is often used in robotics, gaming, and navigation software.

Comprehending Automated Video Redaction

Body-worn cameras are on the rise and are often now a requirement for police officers and their departments. The camera systems are designed for both the officer’s and the public’s safety; when dealing with crime or serious situations it can record any actions that are made against the officers but can also record any questionable police actions or tactics that could be harmful or illegal. News outlets and the public often requisition the footage from these crimes through the freedom of information act, especially after highly publicized cases.

Video redaction machine learning software and its algorithms make it easier to meet that demand. The issue arises when attempting to meet the public’s demand for immediate information while upholding state and local laws regarding personal privacy and sensitive information. An example of this would be displaying minors’ faces, who do not have the adult capacity to give consent and the ability to protect their privacy.

When departments receive requests for this type of information it takes both time and resources while someone is assigned the duty of redacting the video. This is the process of blacking out or blurring minors’ identifying information or other sensitive material. Another issue facing departments is that every state or jurisdiction has its own set of laws regarding how to handle this type of complex data. With many departments facing budget cutbacks, it is difficult to sustain the cost of intensive labor involved for personnel as well as video storage and redaction technology software.

Automated video redaction software helps solve some of the issues for many police departments by eliminating some of the costs involved in maintaining a large personnel department simply to maintain, store, and redact video. Machine learning has advanced to the point that algorithms are now capable of redacting video from cameras in motion, such as the body cams worn by officers. This allows the department to work more efficiently and cost-effectively. After choosing the types of objects or labels to redact from the videos, software programs will now go through frame by frame and locate these targets to blur or black them out accordingly.

While much can be done while the officers are on the job, automated redaction will always still need a human element. Especially in high-stakes cases, if one frame is left unchecked or unredacted, a minor’s face, an important witness or even an undercover officer’s identity can be put at risk. With the possibility of liability on the department, it is important to take all the necessary steps to ensure the public’s safety. A mistake can put someone’s life in danger.

Machine Learning Improves Automated Video Redaction

Since the redaction process is very time and labor-intensive, creating a great expense for most police agencies, automated video redaction helps save time and money. The power of the algorithms now used in machine learning has greatly improved automated video redaction to the point that it is easy to use and effectively relieves some of the costs involved in obscuring personally-identifying information. Generally, there are three key steps in the methods of machine learning used in automated video redaction software. These are:

  • Localization of objects or labels to be redacted.
  • Tracking motion of objects of labels over time.
  • Concealment or obfuscation.

All of these steps can be performed manually, but at great expense to the department. Currently, the best approach is to use a semi-automated method. By using powerful machine learning algorithms and automated video redaction software labels and object information can be determined and blurred or removed as required automatically. However, a manual is necessary to determine the quality of the results. The manual review is important because one frame missed can expose someone’s identity, which defeats the entire purpose and could possibly cost lives.

Powerful Algorithms in Machine Learning Improves Safety for Everyone

While algorithms have become more precise and powerful giving the ability to allow machines to learn and make decisions on their own, there is still a need for human intervention. An example would be when using software to blur or obscure faces in a video, many algorithmic software tools offer an automated face detection selection. However, many can fail to target or locate all faces when dealing with occlusions, side views, or other anomalies. Since lives and criminal cases can be on the line when police are using this software, before allowing any video out for public consumption, a manual review is standard procedure. These reviews catch any missed areas of the film that need to be addressed. The power of these tools in the right hands can help retain information to help solve cases and save lives.