35. Drone Watch: A Novel Dataset for Violent Action Recognition from Aerial Videos

In recent developments, a lot has been done for computer vision applied to human action recognition and violence detection. Although various datasets are available for action and violence recognition, there is a clear lack of datasets that include non-violent and violent activities simultaneously from an aerial view. A new aerial video dataset for concurrent human action recognition, including violence detection, is presented in this study. It consists of 60 min of fully annotated data with two action classes, namely violent and normal (non-violent). The current dataset addresses various factors that are not considered in the existing datasets, like changes in the altitude of the drone, changes in the angle at which the video is being captured, video captured during motion, changes in frame rates, videos from different cameras with different configurations, multiple labels for every subject, and labels for violent activities. The resulting dataset is a multifaceted representation of the real-world scenarios, which addresses various shortfalls in the existing datasets. The current dataset will push forward computer vision applications for action recognition, particularly automated violence detection in real-time video streams from an aerial view. Furthermore, the curated dataset is validated for violence detection using machine and deep learning algorithms, namely Support Vector Machine (SVM), Long Short-Term Memory (LSTM), Bi-Directional LSTM (Bi-LSTM) and Adaptive Boosting (AdaBoost).