Activity recognition algorithms background

The history of recognition with the help of neural networks began in the second half of XX century, but it dates as far back as XIX century, when the scientist Santiago Ramon y Cajal first described the structure of nerve cells, and his followers tried to reproduce this structure. Read more on that in the article about technologies at the heart of face recognition.

Throughout the technology developement, it fulfilled a massive pool of new tasks, such as: recognizing more faces in the frame, recognizing more accurately, recognizing faster. By 2015, the global level of facial biometrics development was quite high. NtechLab has put Russia in the leading position, demonstrating better results than Google and other well-known companies in the international competition of algorithms.

Now that the face recognition accuracy (up to 99%) and speed (less than 0.5 seconds) are so high, it would be appropriate to wonder how to develop the technology further. The battle for fractions of seconds and tiny increments in accuracy percentage will surely rage on, but the existing results already allow to solve the majority of complex tasks with billion-entry datasets, covering all areas of application, both for public safety, and business security. So what’s next?

Recognition technology development trends

Our engineers started their search for the answer a few years ago, which resulted in the creation of silhouette recognition algorithms. Tracking by silhouette — which is also a unique set of human features — allows to detect and instantly count almost any number of people in the tracking zone, even with their backs turned to the camera. It is also capable of tracking their routes within the city based on data from different surveillance cameras.

In 2018, the FindFace solution made it into the top three of the Amazon Pedestrian and Cyclist Detection Competition. Shortly afterwards, NtechLab takes part in an international activity detection competition, where it recognizes, with exceptional precision, 17 different activities in a video stream, such as a phone conversation, typing a text message, leaving a vehicle, moving a heavy object and others, taking the second place.

As part of the contest, participants were required to track the beginning and end of the activity within the raw video stream and send a notification to the organizers. According to Artyom Kuharenko, the architecture of the convolutional neural network developed by the engineers is such that, after the necessary training, the algorithm is able to recognize almost any activity in the video stream.

Evidently, the NtechLab strategists see the further development of identification algorithms by means of neural networks in possibility not only to recognize the person, but also to track his routes of movement, to estimate activities from the point of view of public safety or other criteria, to anticipate dangerous situations and to eliminate them.

In other words, not only the fact of recognition itself is of great value, but also video analytics, with the help of which it is possible to solve various tasks useful for business and the state, and society as a whole.

Activity recognition application scenarios

In regard to the most obvious scenarios of action recognition application, it is worth mentioning the importance of public security threat prevention like tracking the origin of conflicts in the crowd. It is equally important to record cases where someone leaves a suspicious object in a public place and send instant notification to the responsible services.

Then there’s the maintenance of public order. The technology application will allow to successfully fight against smoking in public spaces or using a mobile phone while driving.

Activity recognition is also an effective method of unwanted content control, the manual tracking of which takes more time, costs more and exposes employees to mental stress.

A sequence of control scenarios can be continued with a healthcare use case: the algorithm will monitor compliance with the rules of patient care or will send immediate notification if a patient in the hospital fell out of bed.

Activity recognition has great prospects of application at industrial enterprises, dangerous production and more. NtechLab is already discussing solution implementation options with various industrial, energy and oil and gas companies.

Thus, activity recognition is indeed the highest stage of development of recognition algorithms using neural networks. This is a genuine breakthrough in the industry: the technology allows to make use of even low-resolution cameras and detect the activities of people whose faces cannot be seen.

At the moment, NtechLab is the only award-winning activity recognition developer in Russia. We will keep you updated on the latest developments of our lab engineers.

Subscribe to our newsletter to receive notifications whenever a new post comes up.