A Traversable Areas Detection Model Based on Deformable Attention in Snowy Environments for Autonomous Driving Car

Detecting traversable areas is crucial for the perception system of autonomous driving. Significant progress has been made in recent years in the field of environment perception for autonomous driving. However, several challenges still exist. One of them is the heavy snowfall region in Japan. Most existing image semantic segmentation methods that perform well in clear weather conditions struggle in snowy environments, despite being fine-tuned with data from snowy environments. In addition, a characteristic of the winter snowy environments is the rapid change in road conditions due to factors such as temperature, snowfall, and snow removal operations. Therefore, the model's ability to adapt to changes in different situations is highly required. It is necessary to detect the traversable areas quickly, accurately, and robustly in snowy environments. This study introduces a traversable areas detection model based on vision transformers and deformable attention. We thoroughly analyze the proposed model and compare it with several publicly available deeplearning based models to evaluate their efficiency in detecting traversable areas in snowy environments. The results demonstrate that the proposed model can detect traversable areas in snowy environments both robustly and accurately in real-time, even though the image style of the test set differs from the training one.