Development of robust detector using the weather deep generative model for outdoor monitoring system.

This paper proposes a methodology for building a robust instance segmentation model that can effectively detect objects on construction sites under various weather conditions. We utilize generative adversarial networks (GAN) to create a dataset of construction site images containing different weather conditions and customized the GAN to generate images that reflect the characteristics of the PTZ camera view and weather conditions, while preserving unique construction site entities. The study highlights the importance of modifying deep learning models to fit the unique environment of the construction site to develop models that can detect objects under various weather conditions and improve safety at construction sites. The proposed methodology includes creating a dataset of construction site images that incorporates various weather conditions and developing an instance segmentation model that can be applied effectively to real construction sites. By training the model on images that reflect different weather conditions, the segmentation performance of YolactEdge improved by 2.5% compared to the baseline model. Our future research includes expanding to more complex visual tasks, such as visual relationship and scene-graph generation, to develop even more diverse deep learning models that can be used in construction sites and further improve safety monitoring. • We propose a training strategy for the robust detector under weather conditions. • Our method generates data reflected with weather features in outdoor site images. • The proposed model considers the perspective of top-view without sky feature. • The method can be universally applied to surveillance systems installed outdoors. [ABSTRACT FROM AUTHOR]