![]() | Only 14 pages are availabe for public view |
Abstract ABSTRACT Smart Video Surveillance (SVS) system has been one of the most active research fields. Its applications are extensive. It may include security, public safety, patient monitoring, human activity understanding, and law enforcement purposes. Video monitoring has become a vital aspect of global society since it aids in reducing crime and promoting safety. The traditional surveillance systems depend mainly on persons monitoring the cameras and reporting any suspicious behaviors. However, human-based surveillance systems are impractical. Smart video surveillance replaces the current manual surveillance with an automatic one. The primary purpose of SVS systems is to turn human-based monitoring systems into fully automated systems. Although developing several methods for SVS systems, it still requires significant effort. Surveillance systems are not limited to specific kinds of applications. They can be used with various applications. Our research is employed in detecting suspicious behaviors, specifically carrying light weapons such as knives and pistols. Surveillance systems are perceived to be the most effective way of reducing violence and suspicious behaviors in public environments, including banks, schools, parks, malls, etc. The presence and spread of surveillance cameras in public and private spaces helped increase the scope of research in automated surveillance solutions to protect those places and reduce crime and violence. The SVS systems face many barriers and challenges, and it also requires robust and efficient algorithms for object detection. Still, few algorithms are available that detect weapons, and most of them serve only one weapon type. Unfortunately, this is not the only obstacle that impedes the spread of the SVS systems for detecting Abstract iii suspicious behavior (e.g., crimes, accidents, and robbery). Still, there is also another obstacle, the nature of the weapons themselves. A weapon is a small object in the frame. It is varied in size and shape. Additionally, the knife as a weapon reflects lights that make the knife’s blades less visible in a video sequence. The light reflection on the knife’s surface and the brightness of the knife’s surface makes the detection process extremely difficult, even impossible. For all the reasons mentioned above, turning the traditional surveillance systems into intelligent or smart ones is challenging. This thesis introduces an SVS technique for detecting suspicious behaviors in public and non-public places. This technique can detect the presence of firearms, knives, or other weaponry tools in a threatening or dangerous manner. It reduces the gap between the accuracy of the suspicious action detection and the response time. This technique tries to deal with the illumination problems and lighting problems that face the surveillance systems. This thesis introduces a real-time weapon detection technique using Deep Learning (DL). It automatically classifies the scene as either normal or abnormal by detecting weapons at the video frames. This technique attempts to solve and avoid the majority of the obstacles found in video surveillance systems. This thesis offers a technique capable of analyzing human behavior, especially abnormal behavior using weapons. It introduces an automatic gun detection approach for surveillance video; this approach is divided into the pistol detection approach and the multi-class guns classifier approach. Detecting weapons from surveillance video images is challenging. The different lighting conditions and the blurring of the images due to camera shake are significant obstacles to detecting weapons. Image enhancement is urgent to offer more visibility to the hidden details and reduce blur and noise in these captured images. This Abstract iv research introduces an enhancing surveillance image technique for weapon detection. It eliminates the blurring problem and enhances brightness under different illumination (low and high) conditions. The proposed technique introduces specific criteria to determine whether the image has blurry and brightness problems. It enhances only the noisy images, saving time and making the technique adaptive and suitable for real-time. The suggested technique introduces two procedures, The Blurring Detection and Enhancement Procedure (BDEp) and the Bright / Dark Handler Procedure (BDHp). The BDHp and the BDEp enhance the detection, especially in the worst conditions (i.e., highest brightness and low lighting) in video surveillance. Because there are not enough datasets, a new dataset is prepared and built. This dataset consists of two versions. The first is KwHH_1, with the Total Images being 19565 Images. The second is KwHH_2 Total Images is 22937 Image. The experimental results indicate that the proposed technique helps detect different weapons based on the deep learning approaches. The suggested technique introduces a multi-class classifier that distinguishes between eleven different gun kinds and various knife types. The proposed technique proved its superiority when compared with the state-of-the-art techniques. |