Generation Of Corner Cases In Infrared Autonomous Driving Dataset With Stable Diffusıon

Abstract

The main priority of autonomous driving is situational awareness. To achieve this, autonomous vehicles are equipped with many sensors such as cameras, radar, and LIDAR. With the developments in deep learning, algorithms working with these sensors show very successful results in the detection of popular classes such as pedestrians, cyclists, and traffic signs. However, detecting unexpected objects not encountered in the scene is also very important for safe autonomous driving. Although there are some studies on detecting such situations, which are called corner cases, in the visible domain, these studies assume that these objects are visible. In real life, corner cases can also be encountered in low light conditions such as fog and darkness where visibility is poor. Therefore models developed with the assumption that these objects are always visible fail to detect them in real life, causing the relevant safety precaution systems not to react early and posing a serious risk to traffic safety. Failure to detect or late detection of such situations, even milliseconds are important, causes accidents. On the other hand, since infrared cameras are sensitive to thermal radiation, they can give healthy input without being affected by weather conditions such as darkness, fog, haze or night where visibility is poor, and they enable detection of such corner cases even from long distances. However, no existing infrared autonomous driving dataset contains corner cases and the detection of corner cases in the infrared domain has not been studied before. Therefore, for the first time, we present a high-quality dataset for autonomous driving by generating corner case scenarios in the infrared domain with stable diffusion. In addition, another novel aspect of this study is that we train a detection model on the dataset we have created and perform the detection of corner cases in the infrared domain and present a baseline performance for the model.