HUMAN MOTION PATTERN DETECTION

Abstract: There is a growing need to have effective surveillance systems that can identify and flag alarming behaviour instantly in the current era. In this study, we propose a new method of recognizing suspicious activities of video data using deep learning algorithms. We employed the DCSASS Dataset which contained videos from thirteen categories of suspicious activity, like abuse, arson, assault, robbery and so on. A mixed architecture involving both ResNet50 and I3D was used because it could handle the temporal and spatial complexities that come with video data. The model is trained to recognize subtle cues concerning suspicious behaviour as it exhibits remarkable training accuracy. By subjecting our model to rigorous evaluation on a separate validation set, it shows encouraging results at about 85% accuracy. To improve the performance of the model further, we consider various strategies such as data augmentation, fine-tuning of hyper parameters as well as ensemble methods. We also try to make our models interpretable by employing techniques such as class activation mapping for better understanding of decision making process.

Keywords: DCSASS Dataset, Deep learning algorithms, ResNet50, I3D.