Deep Learning Based Crime Detection from Surveillance Videos

Abstract

The primary objective of deep learning-based crime detection is enhancing public security through active identification of criminal activities. Traditional surveillance systems rely on human operation of live video streams manually, which is time-consuming and prone to human error. To address this, new algorithms such as YOLOv8n and YOLOv11n have been designed as efficient crime detection automation with enhanced accuracy and speed. YOLOv8n and YOLOv11n, both object detection-capable, are particularly trained to rummage through massive databases of surveillance footage and flag abnormal activity or behavior. With the capability to learn patterns and signatures from labeled crime-related data, the models possess the potential to identify potential threats with high levels of accuracy. The performance of the models is measured by metrics like Mean Average Precision (mAP), Precision, and Recall. YOLOv8n scores 0.993 in, precision, 0.993 mAP. Meanwhile, YOLOv11n is an improvement on its previous model, with a recorded precision 0.995, mAP of 0.995 and better accuracy as well as faster processing for criminal detection activities. Finally, the integration of YOLOv8n and YOLOv11n in crime detection systems will revolutionize public security through the identification and prevention of crime. These deep learning models significantly enhance surveillance and provide security personnel with an effective tool to manage criminal behavior. However, concerns regarding data availability, computation power, and ethics need to be addressed to ensure effective and ethical use of this technology.