Detecting Crime Activities Based on Human Behavior Using Machine Learning

Abstract

Object detection is a key concept in optical-based video surveillance software/model/device and automated security system for analyzing/detecting/ identifying crime scene evidence. Photos and video footage can have an important role in the detection of criminal activity. As a massive amount of photos and video footage are collected as visual documentation or evidence of a crime scene the investigation process can be highly complex and may need an advanced technological process. In this research, we have developed a YOLO (You only look once) CNN (Convolution neural network) based real-time object detection model which can automatically detect criminal activities without human instruction. Our model can detect human movements and poses from an image and video footage and classify them into 5 different classes. By analyzing the human poses and movements our goal was to detect and classify the dangerous and suspicious human movements from a crime scene. We trained our model with 1300+ custom images and 5 classes of objects on the google colab with free GPU. We gain an average accuracy of 89% at 0.013 confidence thresholds after training on google colab.