Data augmentation based on deep learning for infrared object detection in autonomous vehicle

Abstract

This paper proposes a data augmentation method for object detection technology that can be used universally in infrared cameras. When acquiring camera-based data, there are cases in which objects cannot be detected entirely due to the limitations of existing visible-light cameras in severe weather or nighttime environments. To overcome these limitations, this study conducted a study on a data augmentation method suitable for deep learning-based object detection of an infrared camera that can be applied in various situations based on the efficiency and characteristics of the infrared camera. In order to obtain research data, image data from infrared cameras and generally visible light cameras were collected in snowy, rainy, and foggy environments, and image learning was performed based on a database suitable for each camera. In the case of object detection of infrared cameras, an adversarial generative neural network was used to generate high-resolution images and follow the distribution of actual samples to overcome the noise and low-resolution limitations. In this paper, 2 data augmentation methods that are optimized for object detection by infrared cameras with the low computational amount and high performance are presented. In particular, to verify the strength and performance of the proposed method, experiments were conducted under various conditions, and the validation of the dataset was completed through qualitative and quantitative evaluation. The proposed methods have the simplicity of high efficiency in terms of computational amount and training time compared to performance, and the flexibility applicable to modules or equipment using infrared cameras. In addition, this study has a high diversity of value for future research through fusion with other sensors. Since the proposed method can be applied irrespective of the type and use of infrared cameras, it can be applied to all fields in various forms, such as autonomous vehicles, military GOP scientific guard system and industrial infrared sensors.