Evaluation of Geometric and Photometric Data Augmentation for Pedestrian Detection with Thermal Cameras

The capability to detect pedestrians using video surveillance systems is a fundamental task to assess the number of people in particular areas and to enable other applications like people tracking and trajectory estimation. Thanks to recent advancements of object detectors based on deep learning approaches, it is possible to localize pedestrians in recorded images with good accuracy. However, some limitations still exist. Most of the proposed solutions at the state-of-the￾art deal with color RGB images, that could present privacy issues, hampering the deployment of such solutions. Moreover, changing the application scenario could worsen the performance, requiring time consuming fine tuning and transfer learning operation. In this paper, we respond to these issues by using thermal images analysis that can guarantee a higher privacy preservation instead of RGB data. In addition, a deep study on data augmentation techniques has been provided to generate more flexible trained models that could simplify their adoption in different scenarios without requiring additional acquisition, labeling work and fine tuning.