A method has been discovered to enhance the accuracy of human action recognition using surveillance cameras.

In strategically important industrial facilities, crowded areas, shopping centers, concert halls, and educational institutions, surveillance cameras are being installed. These cameras are equipped with human motion recognition technology that detects and classifies moving objects within their field of view. Based on the nature of the actions and established restrictions, they generate a corresponding response. In the event of a threat, preventive measures against emergencies can be initiated. However, the effective operation of such a system and timely responses rely on the accuracy and speed of situational recognition. Researchers from Perm Polytechnic University have proposed a mathematical model for describing human behavior, which enhances the accuracy of image recognition from surveillance cameras to 95 percent.

2024-12-19 10:00:35https://naked-science.ru/article/column/dejstvij-cheloveka-s-kame

The article was published in the journal "Herald of PNIPU. Electrical Engineering, Information Technologies, Control Systems." The research was conducted as part of the "Priority 2030" program.

To recognize actions using surveillance cameras, it is essential to identify a person as a distinct object and gather information about their body position and movement sequence. Additionally, these data must be stored for further processing and classification tasks. A mathematical model and its representation play a crucial role in this process.

Frame-based description models, where actions in each frame are identified separately, have a drawback: other people and objects may be present in the field, leading to incorrect analysis of the information. To mitigate this issue, a vector model is often utilized—movements are determined through a sequence of coordinates of key points on the human skeleton. To enhance accuracy, these points are grouped, enabling the algorithm to find and process information about various body parts.

For precise object recognition in space, data undergo normalization, converting key points from pixel values to real-world measurements. Existing methods do not account for the significant variability in the person's orientation and position in space. The same movements, performed with different offsets relative to the camera, are likely to be recognized as different actions. This often complicates system functionality, requiring a substantial increase in device memory and complicating calculation algorithms, which may not always be practically feasible in terms of time and financial resources.

Researchers at Perm Polytechnic University have discovered a way to speed up the processing of video material and improve the accuracy of movement detection. They analyzed existing human skeleton recognition models and processing algorithms. Based on their findings, they proposed the implementation of an original model and video image normalization technology in surveillance systems.

“We created a simplified model that excludes unnecessary information for our research, such as the position of hand fingers. Often, their locations are noisy, which also consumes processing time and complicates action recognition. The key points in our model are the eyes, shoulders, hips, elbows, wrists, knees, and feet. We also proposed an algorithm to transform information about human skeleton movement, recognizing actions by comparing data from different cameras or angles for greater accuracy,” says Alexander Knyazev, a graduate student at the Department of Information Technologies and Automated Systems at PNIPU.

“Experiments showed that our model and video image normalization technology achieved a recognition accuracy of 95 percent. In contrast, the use of primary data only provided 35 percent accuracy,” comments Rustam Faizrakhamanov, head of the Department of Information Technologies and Automated Systems at PNIPU, Doctor of Economic Sciences.

The implementation of the development by Perm Polytechnic University researchers will enhance the accuracy of human action recognition from surveillance cameras, which is effective for monitoring and ensuring safety in industrial enterprises, secured areas, and public places. Several industrial companies have already shown interest in the developed technology. The initiative is supported by the Innovation Promotion Fund, with a grant approved under the Start-1 program.