Anomaly detection and size reduction system with deep learning approaches in videos

In institutions and organizations, security camera systems are usually used to detect abnormal situations in videos. Security personnel resources are used in the management of security camera systems. Thanks to the attention and follow-up of the security personnel, an abnormal situation is detected in the video images on the security camera systems and the necessary interventions are made. Due to the fact that the source of security personnel is human, many human structures such as physiological, biological and spiritual affect the attention and monitoring they should have. In addition to instant monitoring of security camera videos, it is also very important to store them. The videos must be recorded and stored in order for detailed analysis operations and legal proceedings to be carried out on the videos. The retrospective temporal increase in the storage of video recordings in security camera systems directly increases the usefulness, reliability of the system and the cost of the system, which is a disadvantageous situation. This thesis study is planned in two stages. In the first stage, it is aimed to detect abnormal situations in videos with computerized imaging techniques based on deep learning and not affected by human structures. As a result of the work at this stage, an automatic weapon, knife and violence detection system was made from abnormal situations in instant video images. Since this study was conducted during the Covid-19 pandemic, the data sets were compiled from anonymous and copyright-free internet environments, YouTube images and academic study data sets with open source access in the model trainings conducted to enable abnormal situation detection operations in videos. YOLOv4, YOLOv5, YOLOX and YOLOR models based on deep learning based image processing techniques were used in experimental studies. The highest performance value was achieved as 97.6% with the YOLOR model. For the detection of violence, EfficientDet D0-D7, YOLOv5 and Faster R-CNN models were applied and the best success rate was achieved with the EfficientDET D0-D7 model at 94.1%. In the second phase of this thesis study, abnormal situation classification in security camera video recording images and size reduction in the storage of images were studied. The data sets for image classification were created from the compilation of anonymous copyright-free data sets, as in the detection and detection of abnormal situations in the first stage of the thesis. For these classification operations, AlexNet, GoogLeNet, VGG16, VGG19, SqueezeNet, Inception, ResNet18 and ResNet50, which are pre-trained deep CNN architectures based on deep learning, and LSTM and GRU from RNN architectures were used. The best performance value in the classification of weapons and knives from abnormal situations was obtained with the VGG16 model, which was trained based on a fine-tuning approach. The best performance value in the violence classification study was obtained with a 100% accuracy score using the AlexNet model. For the size reduction study, an original size reduction model was produced by calculating the classification data and basic image processing techniques and video picture frame differences made in the second stage of the thesis. This model has been developed taking into account details such as the ability of users to use it easily, checking its sensitivity according to the requirements of institutions and organizations, and the ability of the system supervisor to provide technical support. This system is designed through the Matlab AppDesigner platform in accordance with human-computer interaction. In this thesis study, experimental studies based on the detection and classification of violence, weapons and knives were carried out. In addition, an original size reduction system has been developed to make these experimental studies feasible and to make gains in the storage of images. As a result of these studies, violence detection, weapon and knife detection and classification studies have been successfully carried out for use in analysis operations in instant and image archive recordings in video systems. Instant detection studies in images can be connected directly to security camera systems and used. The archival image analysis studies carried out were connected to the image analysis and size reduction system prepared by observing the rules of human computer interactions specific to this study. In this study, it has been observed that it provides a gain between 10% and 40% in stationary environments thanks to the original Decimation system. As a result, all the goals and objectives have been successfully fulfilled in this thesis study.