Video analytics using deep learning for hajj pilgrimage crowd monitoring
This research advances crowd analysis and surveillance with a unique interest in excessive crowd density, and abnormal crowd detection for Hajj and Umrah pilgrimages. Gathering a large number of people in a shared physical area is very common in urban culture. Although there are limitless examples o...
Saved in:
Main Author: | |
---|---|
Format: | Thesis |
Published: |
2022
|
Subjects: | |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Summary: | This research advances crowd analysis and surveillance with a unique interest in excessive crowd density, and abnormal crowd detection for Hajj and Umrah pilgrimages. Gathering a large number of people in a shared physical area is very common in urban culture. Although there are limitless examples of mega crowds, the Islamic religious ritual, the Hajj and Umrah, is considered as one of the greatest crowd scenarios in the world. The Hajj is carried out once in a year with a congregation of millions of people when the Muslims visit the holy city of mecca at a given time and date. Such a big crowd is always prone to public safety issues, and therefore requires proper measures to ensure safe and comfortable arrangement. Crowd surveillance has suffered numerous conditions, including distinctive scenes of crowd density. This thesis has three main contributions which include crowd density estimation using deep learning, crowd anomaly detection using deep learning and development of the crowd dataset with augmentation techniques. Firstly, this thesis aims to address current technological challenges faced in video analysis in a scenario where the movement of large numbers of pilgrims with density ranging between 7 to 8 per square meter. In current crowd density estimation video analysis, less number of crowds particularly for Hajj, manual and semiautomatic techniques are employed. This research proposes a fully automatic approach to overcome the problem of estimating crowd density taken by a surveillance camera at a distance. The proposed approach employs a fully convolutional neural network (FCNN)-based method to monitor crowds, especially for the classification of crowd density. For this experiment, we have used an FCNN-based method. Because FCNN is good for crowd analysis and has better accuracy, Currently, there are a few traditional models such as ResNet, VGGNet, AlaxNet, CNN, etc. The FCNN is a very fast and efficient model. For big crowd analysis undoubtedly FCNN is the best model in this domain. To validate the proposed method, we compared with existing models using existing datasets. The proposed FCNN based method achieved a final accuracy of 100%, 98%, and 98.16% on the new Hajj-Crowd-2021 dataset, the UCSD dataset, and the JHU-CROWD dataset, respectively. Additionally, The ResNet based method obtained final accuracy of 97%, 89%, and 97% for the new Hajj-Crowd-2021 dataset, UCSD dataset and JHU-CROWD dataset, respectively. The proposed Hajj-Crowd2021 crowd analysis dataset and the model outperformed the other state-of-the-art datasets and models in most cases. Secondly, this architecture enables us, through spatial-temporal convolutions, to capture features in both spatial and time dimensions and to extract knowledge about the presence and motion encrypted in continuous frames. Two primary goals are addressed by this FCNN-based architecture: feature representation and wrong movement outlier identification. The suggested technique exceeds current methods in terms of detection and classification accuracy, as shown by experimental results on benchmark datasets. The empirical findings show that our model can deal with various scenarios and identify anomalies. We employ the widely known datasets of UCSD and Subway and present the Hajj-Crowd-2021 as a new video dataset. We proposed the Crowd Anomaly Hajj Monitor (CAHM) method using video sequences of crowded scenes to identify and localize wrong movement abnormal behavior. The main contribution of our approach is the combination of anomaly detection-based optical flow features and classification based on spatial-temporal features using fully convolutional neural networks (FCNN), which has never been achieved in the past, based on our extensive research in this domain. In this research, we detected the anomaly from the video by using optical flow (the Lucas-Kandae method). For detecting the anomaly using optical flow it is very easy and accurate. Using fully convolutional neural networks (FCNN) and spatial-temporal data, a pre-trained supervised FCNN detects anomalies in Hajj crowded scenes. To verify our work, we used the available datasets to compare the proposed model to current models. In all of these situations, our model outperforms. Crowd analysis, for instance, improves classification accuracy by achieving 96% in highly crowded crowd situations. Furthermore, crowd anomaly analysis improves classification accuracy for the UCSD, Ped1, and Ped2 datasets for extremely congested crowd situations by 89%, 83%, and 85%, respectively. On the other hand, our method has improved classification accuracy for the Subway dataset, 88% for the exit, and 85% for the Entrance case. Thirdly, the majority of crowd datasets have very few training samples, while algorithms based on deep learning need large amounts of training data. In this thesis, we introduce a deep Hajj crowd dilated convolutional neural network (DHCDCNNet) for crowd density analysis. This research also presents augmentation techniques to create additional datasets based on the Hajj pilgrimage scenario. The majority of prior research has employed deep and shallow networks, two distinct kinds of networks, to extract high and low-level features. We utilize a single framework to extract both highlevel and low-level features. For creating additional datasets we divide the process of image augmentation into two routes. Each route consists of different image processing modules. In the first route, we utilize magnitude extraction followed by the polar magnitude. In the second route, we perform a morphological operation followed by transforming the image into a skeleton. This thesis presented a solution to the challenge of measuring crowd density using a surveillance camera pointed at a distance. An FCNN-based technique for crowd analysis is included in the proposed methodology, particularly for classifying crowd density. With the augmented dataset the FCNN method has achieved accuracy of 97% , 89% and 100% for the JHU-CROWD dataset, the UCSD dataset and proposed Hajj-Crowd dataset respectively. With the VGGNet approach, we achieved accuracy of 98 %, 97 %, and 97 % for the new Hajj-Crowd2021 dataset, the UCSD dataset, and the JHU-CROW dataset respectively. Using the ResNet50 approach, the proposed Hajj-Crowd dataset, the UCSD dataset, and the JHU-CROW dataset all had an accuracy of 99%, 91%, and 97%, respectively. As a result, DHCDCNNet outperformed current techniques in both sparse and dense crowd scenarios. Finally, this work attempts to address the above challenges by developing a new Hajj-Crowd-2021 crowd image dataset based on the Hajj pilgrimage scenario. This research also intends to develop another new Hajj-Crowd-2021 anomaly video dataset based on the Hajj pilgrimage scenario in order to solve the mentioned issues. The Hajj crowd dataset contains a total of 30,000 images. The Hajj crowd video anomaly dataset consists of a total of 100 anomalous and 100 normal videos, respectively. |
---|