Person Activity Classification from an Aerial Sensor Based on a Multi-level Deep Features

The intelligent surveillance system is considered as an important research field since it provides continuous personal security solution. In fact, this surveillance system support security guards by warning them in suspicious situations such as recognize the abnormal person activity. In this context, we introduced a new method for person activity classification that includes offline and inference phases. Based on convolutional neural networks, the offline phase aims to generate the person activity model. Whereas, the inference phase relies on the generate model to classify the person's activity. The main contribution of the proposed method is to introduce a multi-level deep features to handle inter- and intra-class variation. Through a comparative study, performed on the UCF-ARG dataset, we assessed the performance of our method compared to the state-of-the-art works.