Human Behavior dataset (HBD21)
HBD21 dataset comprised of video files. This video dataset is divided into four categories i) Assault violence, ii) Gun violence, iii) Sabotage violence, and iv) Normal actions. This dataset consists of 463 videos. Each category has contained more than 100 video files. Each video's length is 9 seconds, and it is observed that this length is sufficient to represent the real action. There are three separate folders attached in HBD21 dataset.rar file. The first folder represents the entire original dataset files, while in the second folder, the train_split and test_split datasets are present. The training and testing video files are in the ratio of 7:3, i.e., 70% training and 30% testing videos of the original dataset. Further, the third folder consists of a pre-processed enhanced version of the training and testing dataset. In this improved version, each video's brightness is increased at a suitable threshold value, which helps in achieving achieve better performance in the training and testing stages. HBD21 dataset is collected from a mobile camera of about 16MP, placed at a tripod with a 5m to 12m distance. The resolution of each video is 640*480. This dataset is captured at various locations in different lighting conditions using multiple objects(bat, badminton, gun, stick, axes, chairs, bottles, and many more).
Steps to reproduce
This dataset is very useful for violence detection or can distinguish the action between unusual and usual activities. Researchers can apply action detection techniques to classify or detect the behaviors of humans. With the help of a mobile camera and tripod stand, the dataset has been captured. The image enhancement technique is used to increase the brightness of the dataset to enhance the results. Researchers can also apply any other preprocessing techniques like augmentation, noise removal, background subtraction, and many more to increase the dataset's size and improve the videos to reproduce the research area. The preprocessed videos obtained by applying these techniques help the models to classify the actions more accurately. This HBD21 dataset is reliable for all types of learning-based approaches.