Convolutional neural network and unmanned aerial vehicle‐based public safety framework for human life protection

Abstract

SummaryIn this paper, we developed an object detection and identification framework to bolster public safety. Before developing the proposed framework, several existing frameworks were analyzed to bolster public safety. The other models were carefully observed for their strengths and weaknesses based on the machine learning and deep learning algorithms they operate on. All these were kept in mind during the development of the proposed model. The proposed framework consists of an unmanned aerial vehicle (UAV) utilized for data collection that constantly monitors and captures the images of the designated areas. A convolutional neural network (CNN) model is developed to recognize a threat and identifies various handheld objects, such as guns and knives, which facilitate criminals to commit crimes. The proposed CNN model comprises 16 layers with input, convolutional, dense, max‐pool, and flattened layers of different dimensions. For that, a benchmarked dataset, that is, small objects handled similarly to a weapon (SOHAs), a weapon detection dataset is used. It comprises six classes of 8945 images, with 5947 used for training, 1699 used for testing, and 849 used for validation. Once the CNN model accomplishes the object identification and classification, that is, the person is criminal or non‐criminal, the criminal is forwarded to various law enforcement agencies and non‐criminal data are again forwarded to the CNN model for improvising its accuracy rate. As a result, the proposed CNN model outperforms several pre‐trained models with an accuracy of 0.8352 and a validation accuracy of 0.7758. In addition, the proposed model gives a minimal loss of 0.83 with a validation loss of 0.97. The proposed framework decreases the burden on crime‐fighting agencies and increases the accuracy of crime detection. Additionally, it ensures fairness and operates at a meager computational cost compared to similar pre‐trained models.