The recognition of dangerous objects is an important safety factor in today’s world. In our work, we focused on creating a large dataset that would contain dangerous objects, mainly guns. We created a dataset with two categories called Guns and People. We created the dataset using other freely available datasets, which we later modified. We removed corrupted, repetitive images from the dataset and later expanded the whole dataset. Experiments were then performed on the created data set. We used freely available neural networks such as ResNet and InceptionV3 on the dataset and then designed our own architecture. The best results were achieved by the InceptionV3 neural network with an accuracy of 96.67%, our proposed architecture achieved 88.70%. However, all tested neural networks achieved very good results, which confirms the reliability of the dataset for training neural networks.