Contextual Crime Scene Interpretation using Artificial Intelligence Techniques

Abstract

Crime prediction is always a challenging task from the surveillance point newlineof view. An essential and crucial component of any criminal investigative newlineprocedure is crime scene interpretation. It acts as a systematic and newlineanalytical way to go down the many components and evidence found in a newlinecrime scene. Its principal aim is to reconstruct the sequence of events that newlinetook place at the crime scene. The combination of technology and learning newlinealgorithms has brought about a tremendous revolution in the field of crime newlinedetection in current times. This novel strategy increases public awareness newlineand aids law enforcement agencies by utilizing state-of-the-art instruments newlineand procedures to detect the presence of firearms at crime scenes or newlinepossible criminal activity. These systems use sophisticated algorithms to newlineanalyze these visual data sources in real time, focusing primarily on the newlinedetection of weapons such as knives and guns in the scene. newlineThis thesis introduces a novel approach to detect weapons and individuals newlinein both static and video scenes. The presence of people involved in newlinequestionable or potentially dangerous activities so the method not only newlineidentifies weapons but also tracks them in relation to individuals, newlinecontributing to the prediction of potential criminal activities. The dataset newlineemployed in this thesis is a custom dataset containing images of three newlineclasses: guns, knives, and persons. Each image is meticulously annotated newlinewith manually marked object positions and object classes. The proposed newlinemodels combine the Faster RCNN and YOLO v4 models. The Faster newlineRCNN is a region-based convolutional neural network that excels in object newline newlinedetection tasks, while YOLO (You Only Look Once) is known for its real- newlinetime object detection capabilities. This combination aims to leverage the newline newlinestrengths of both models for enhanced performance. In addition to newlinetraditional object detection techniques, the thesis incorporates the use of newline newlinevii newline newlineMediaPipe functions to generate data points on the human body. These data newlinepoints play a crucial role i