Real-Time Crime Detection System for Public Safety Based on YOLO Model and Transfer Learning

Abstract

Crime is a global threat that can affect individuals and any nation when it is not adequately curtailed. The literature survey, global news, and reports in Nigeria show that security threats and crime rates have heightened the need for efficient techniques for crime detection. Earlier studies have explored crime prevention using machine learning models such as k-Nearest Neighbors (kNN) and support vector machines (SVM) to forecast crime. However, considering the dynamic nature and changing patterns in crime trends pose limitations in these methods. Conversely, the Convolutional Neural Network (CNN)-based deep learning models has shown significant potential in this regard. The You Only Look Once (YOLO) object detection, an example of a CNN model, holds promise for real-time detection. This study proposes a real-time crime detection system based on the YOLO model and transfer learning. An algorithm was formulated based on YOLOv5 architecture and was implemented in Python using PyTorch deep learning framework in a Google Colab environment. The model was trained on a crime custom dataset that contains different classes of weapons and violent crime classes. Transfer learning was employed to facilitate the training. The model was then evaluated using the mean average precision (mAP) metric and F1 score, yielding a promising result of approximately 0.81 and 0.80, and a speed of 94 Frames Per Second (FPS). To test for real-time inference, the model was integrated into a user-friendly interface that was developed using Streamlit and OpenCV, and real-time testing was performed. The results demonstrate the system's ability to detect crime events in real-time. Findings from the study showed that the proposed system when deployed holds significant potential to enhance public safety and support security agencies' efforts in curbing crime.