Semantic Segmentation Research of Motion Blurred Images by Event Camera

The semantic segmentation of road objects is a prerequisite for autonomous driving. Rapid or bumpy movement of vehicles can lead to the images to be blurred, which reduces the safety of autonomous driving. This study combines the dynamic visual sensor (DVS) and RGB camera to improve the detection accuracy of road semantic segmentation. Firstly, based on DVS and RGB images, we develop two datasets of blurred images, including the stone and run datasets. Then, a double channels’ detection network is designed with several Swin-transformer modules, and the two channels are designed to extract features from DVS and RGB images, respectively. In addition, a feature fusion module is designed to extract attention from DVS images. Finally, experiments demonstrate that our proposed double-channels’ fusion method achieves the state-of-the-art segmentation performance for blurred images.