A chess computer, self-driving cars, Siri and Google answering spoken questions, spam filters, and even a calculator solving a math problem for you—all of these help us in our daily lives and make use of machine learning. Machine learning, a form of artificial intelligence (AI), focuses on designing machines/devices that can learn from data. They do this with the help of algorithms programmed by humans.
Recognizing movement
This is not simple. To begin with, we must ‘tell’ the software that a certain group of changing pixels in the image represents a moving human being. And that humans can move in different ways, such as crawling, walking, or cycling. After this, the software is capable of independently—using the correct algorithm—filling in the infinite number of possible positions between walking and crawling. Innocuous, repetitive movements, like trees, bushes, flags, and noise in the image, can also be filtered by intelligent software. It is possible to instruct the software to ignore the part of the image where the bush is moving. However, this also means that no detection will take place in the area where the bush is moving.
Patterns in movement
The next step is to teach the software how to distinguish an intruder from an innocent person. Finding suspicious individuals in recorded footage is easier if we know the time of the break-in. And by having the software operate during times when the presence of people on the premises is unusual, we can go a long way in detecting unwanted activities in time. But we also want to recognize (suspicious) behavior. Patterns in movement can be analyzed. For this, we can use specific algorithms, such as loitering and trip wires. These algorithms are designed, programmed, and now widely available thanks to human effort. In the future, more and better algorithms will continue to be developed, contributing to more accurate recognition.
Humans provide the intelligence
The intelligence in the camera system (i.e., the algorithms) still comes from us, the humans. By adding more of our intelligence to the software, it will gradually become smarter. This learning process takes time. At the moment, only relatively simple intelligence is available. Is it a person, a tree, or a car? Is the car moving or stationary? Are people standing still or running? The perfect algorithm for detecting suspicious behavior doesn’t exist yet. It is challenging to develop a single algorithm that can detect all forms of suspicious behavior because it strongly depends on the local situation. Therefore, customization is required.
Software processes data faster than humans
Humans are still smarter and more flexible than even the most intelligent software, especially when it comes to recognizing and assessing situations. The strength of intelligent software lies primarily in the speed with which it can process vast amounts of information (data). And software is never distracted or tired. Self-driving cars still make mistakes because the learning process is ongoing, but ultimately, the car will be able to respond faster than we can to a dangerous situation—even while simultaneously finding us a better radio station.
The future
In the end, the development of hardware and intelligent software will help us more efficiently analyze behavior and find the right camera footage. Machine learning is in progress…