Panic detection by regional velocity changes in crowded areas from surveillance vi̇deo

The automatic extraction of information from the surveillance camera has become a matter of frequent work with the widespread use of surveillance cameras. City surveillance cameras monitored from a single coordination center by authorized personnel. Especially in situations where immediate intervention is required, the presence of a human factor can lead to faults and delays in determining the situation. In areas with a high concentration of human, it is vital that reaching of emergency aid team at the appropriate location immediately after a suspicious situation has occurred. In this study, a system that detects panic situation from surveillance camera is designed and realized. When a suspicious situation occurs, there may be a velocity increase in situations such as running away, or a velocity decrease in situations such as an injury. Panic states were determined by evaluating these changes at the velocity of people. People in the field of view of the camera have been identified using CNN (Convolutional Neural Network) for the realization of the panic detection system. Then, from the previous locations of these people, the possible locations are calculated by the Kalman filter. From the calculated positions, the most suitable one in terms of appearance properties was selected by the Hungarian algorithm. The use of the Hungarian algorithm with the CNN makes it possible to find people who have been interrupted by staying behind other objects. Orbits were extracted with a short follow-up of people, and the average speeds of each person appearing on the screen were calculated. When the whole image is evaluated, it is perceived that the movements of persons close to the camera are faster than those who are away from the camera. By separating the images into regions, the velocity changes of the people in the regions have been evaluated within themselves. At this point, the speed misalignment which is formed in perspective from the viewpoint is eliminated and the contribution to panic detection is investigated. It has been observed that differences in velocity changes in panic and non-panic regions increase false positive detections. The combination of seperating image into regions and adaptive threshold value methods reduced the false positives, so that unnecessary warning situations are minimized in the coordination center.