تعداد وتقدير كثافة الحشود اعتماداً على الشبكات العصبية الالتفافية لمشاهد الحج والعمرة

Abstract

Crowd counting or crowd estimation is a method used to count or estimate the number of people in an image/video. Visually analyzing crowds, crowd counting, and density estimation are challenging yet exciting tasks in the computer vision field and can enhance several crowd related tasks regarding safety and security. Intelligent crowd counting refers to innovative methods like adaptive crowd control and large-scale surveillance by integrating Machine Learning (ML) and Artificial Intelligence (AI) technologies with traditional crowd counting techniques. Many existing approaches count people in a single frame using techniques based on detection, tracking, regression, and activity recognition. However, they are only applicable to sparse scenes, and their efficiency drastically decreases when applied to crowd scenes that contain a large number of people with frequent and heavy occlusions. Consequently, many new research projects, particularly those focusing on crowded scenes, have been conducted recently. In particular, Convolutional Neural Networks (CNNs) have proven to be helpful and more accurate in predicting crowd counts and have been tested with various challenging datasets; it is also one of the most powerful deep learning models that have proven effective in many areas. Excellent performance has been achieved by CNN and overcomes the limitations of traditional crowd counting methods.

In this thesis, a new dataset of Hajj and Umrah scenes has been built to count the crowds during Islamic rites' performance. Hence, this research focuses on counting the number of people in Hajj and Umrah scenes using CNN architectures on a custom dataset. A CSRNet (Congested Scenes Recognition Net) model was used to take advantage of CNN techniques. The CSRNet is a deep CNN architecture used to generate high-quality density maps and get an accurate estimation of the number of crowds from the Hajj and Umrah scenes in Makkah after collecting and annotating our dataset for this purpose. The performance of the CSRNet model was compared with a Bayesian-loss model, and satisfactory results were obtained on our curated dataset. We also provided an overview of the uses of crowd counting traditional methods, which are handcrafted feature extraction methods. We provide a detailed overview of crowd counting using Convolutional Neural Network (CNN) techniques. The thesis pro vides an extensive literature review specifically for crowd counting using the CNN method because it is considered the most powerful crowd counting and analysis technique, nowadays. We also discuss related challenges, such as complex backgrounds, perspective variation, objects clutter, for which CNN has proven its effectiveness. This research provides a robust and easy-to-understand background to crowd counting and the latest research findings in the field.