الفهرس 
AbstractOnly 14 pages are availabe for public view
 |  | from | 104 |  |  |
| | | from | 104 | | |
Abstract
The crowd phenomenon has become of great importance because of the growing number of popular events that bring together large numbers of people, such as markets, subways, religious celebrations, and sporting events. Now they are confronted with multiple security issues. Therefore, the necessity for representation of analysing and understanding the behaviour of human crowds is gaining importance, especially in the context of citizen security, safety, and city administration following the ubiquity of video surveillance. This is done to detect potentially unsafe situations and avoid excessive congestion. Despite the number of studies done in this field, understanding the dynamics of crowds still has many challenges. A crowd can quickly grow out of control, making it difficult for organizers to maintain control. Problems, such as abnormal behavior or even stampedes, could emerge in such situations. Therefore, it is critical to have an intelligent crowd analysis system in place to ensure public safety and maintain high pedestrian flow throughput to avoid stampedes. Another challenge is obtaining the correct foreground pixels that exclusively show people. Furthermore, strong feature extraction approaches and powerful and reliable decision-making classifiers are required for human crowd behavior systems to perceive and interpret crowd behavior effectively. This thesis aims to improve crowd analysis approaches for detecting and identifying human activity in crowded videos. A carefully engineered pipeline for crowd analysis is proposed, integrating the capability of deep neural network (DNN) algorithms. The pipeline includes three principal stages that cover crowd analysis challenges. First, human detection is performed using the You Only Look Once (YOLO) model for object detection and the Simple Online and Real-time Tracking (SORT) algorithm. The SORT method uses a combination of the Kalman filter and the Hungarian algorithm to detect and track target objects in real-time video frames. Second, a specific location’s density map and crowd counting are generated using bounding boxes from an object detector. Finally, in order to classify normal or abnormal crowds, individual activities are identified with pose estimation. An action classifier is used to classify activities to determine abnormal activities in the crowd.