Criminal activity detection using deep learning algorithms

Abstract

Criminal activities using guns and knives occur very frequently. The quick and accurate detection of a criminal activity is paramount to securing a place where people usually gather every day. More and more security systems are being developed as the number of cities are growing rapidly. This creates a backlog of video data that is being monitored under human supervision but usually human error happens in such cases. This also creates a huge amount of workload for the supervising team. There are several solutions in computer science that can be implemented for immediate and accurate criminal activity detection without any human intervention. Human behavior and pattern recognition is a challenge when it comes to criminal detection as there are several people who act in an abnormal way but aren’t suspicious. In such cases it might generate a false alarm. As we proceed further with our research, most of the crimes take place with the use of handguns or knives. There are many more studies from di↵erent countries that show that, most dangerous crimes took place using weapons of di↵erent sorts. So, in order to detect a criminal from a live crime scene, the first and the quickest step is to determine whether a person is carrying an arm or not. For such detection method, Convolutional Neural Networks is very useful. That’s why among all di↵erent types of Deep Learning approaches, we opted for Convolutional Neural Network (CNN) to identify a criminal using object detection method. The major challenge of our research was the unavailability of datasets. We created our own image dataset and classified them into four different classes in order to train our model. We have 4,180 images in our dataset which are collected from di↵erent crime scenes. There are several CNN models that give efficient results in terms of object detection from image datasets. In our work, we implemented five di↵erent CNN models which are MobileNetV2, Inception-v3, Xception, VGG16, ResNet50 and as a result accuracy for each model is 98%, 98%, 94%, 70% and 60% respectively. The accuracy in MobileNetV2 and Inception-v3 was the highest.