Combination of transfer learning and incremental learning for anomalous activity detection in videos
Yash Rathore, Vandana Japtap, Parth Gawande, Sarthak Oke, Aditya Rao · 2023
Human action recognition has turned into an excellent exploration region in computer vision issues. There are numerous applications where human activity detection can be consolidated, for example, analysing the video information which will assist with observing information produced by CCTV cameras. The main drawback of training video data is that assuming new classes should be instructed to the framework, the system should be retrained without any training, and all classes retaught to the system. This particular problem has many difficulties, for example, putting away and holding information and responding training costs hence we use incremental learning to avoid catastrophic forgetting which is learning new examples without compromising previously learned knowledge. There is a requirement for transfer learning algorithms because existing models are reused to take care of another test or issue and to save time and resources from being required to prepare different machine learning models to complete similar tasks while at the equivalent time upgrading their overall performance. In this paper, a system is proposed which uses transfer learning along with incremental learning to detect differ- ent types of human activities in videos. The proposed system is evaluated on a benchmark dataset, KTH, UCF101 (UCF101 human actions dataset. experiments performed on complicated video datasets demonstrate the ability of the proposed method for low complexity incremental learning while achieving significantly better accuracy than existing such models.