Rotation Aware Object Detection Model with Applications to Weapons Spotting in Surveillance Videos

Abstract

Detection of weapons automatically is very important for improving the security and prosperity of people, in any case, it is a troublesome undertaking due to huge assortment of size, shape and presence of weapons. View point varieties what’s more, impediment likewise are the reasons which makes this errand more troublesome. Further, the present detection algorithms of objects process rectangular areas, anyway a thin and long rifle may truly cover simply a little part of zone and the rest may contain unessential subtleties. To beat this issue, we propose a deep learning based model for detection of weapons with orientation, which not only gives rotation aware bound box but also improves the detection performance. The proposed model provides orientation with the help of angle classification by dividing angle into eight different classes. To train our model for weapon recognition another new dataset containing of around 6400 pictures is assembled from the web and afterward manually annotated that. We also provide three standard horizontal annotation format of our dataset as ground truth along with oriented ground truth for further exploration in future. The proposed model is assessed on this dataset, also, the near investigation with off-the rack object indicators yields predominant execution of proposed model, estimated with the standard assessment procedures. The dataset and the model implementation are made publicly available at this link: https://bit.ly/2TyZICF.