An application towards detecting criminal behaviors in indoor spaces with deep learning

Due to their nature, crime and criminal events have always been important for human and social life in the past, and it is clearly observed that they would have important effects in the future. In addition to legal sanctions, it is possible to say that limited technological applications reduce the crime-related incidents to a certain extent, however, they are far from meeting the expected requirements. One of the most important reasons for this is that criminal events are examined asynchronously and manually after they occur, and technological tools are often used for recording and processing the data belonging to these events. Therefore, it is considered that smart technological solutions that would monitor the events in real time and automatically analyze the criminal behaviors may reduce crimes and also increase the deterrence for crime. The main purpose of the study is, therefore, to produce a smart solution for detecting criminal behavior in real time by using video images obtained from security cameras (CCTV) in closed areas. An integrated research is carried out using deep learning methods, techniques and tools within the framework of design science research method. The theoretical foundations of the research are crime and crime detection methods, artificial intelligence, artificial neural networks, deep learning and image processing knowledge areas. The proposed deep learning solution and its original models and algorithms constitute the main outputs of this study. Additionally, criminal behaviors have been determined for the solution of the research problem. These are fighting, running, aiming a gun, jumping, raising hands, lying on the ground, inaction positions and others. Security Camera Crime Data Set consisting of 617 real video images taken from security cameras has also been created. As a result of the study, it is found that unsupervised learning models are more effective in crime detection than supervised learning models. In line with the findings and research observations, it is possible to say that the solution, developed using deep learning methods, meets the aims and objectives of the research. However, considering the limitations of the research, it should be noted that there are some aspects that need improvement with future studies.