Image Reconstruction Leveraging Time-Event Feature Fusion for Event Cameras

The event cameras generate asynchronous event sequences, rendering most existing image-based algorithms in-applicable for direct use and processing. Therefore, the development of a simulator that utilizes event streams captured by an event camera as input and reconstructs image sequences as output is crucial for subsequent tasks such as high-speed moving object tracking, object recognition and detection, and SLAM. Despite the extensive research and attention garnered by deep learning-based image sequence generation, there remains a problem of local representation information loss during the reconstruction process from event streams to image sequences. Existing solutions, whether based on Stacking Based on Time (SBT) or Stacking Based on Events (SBE), exhibit significant limitations when using a single stacking approach. To address the insufficiency in feature representation capability, this paper proposes a visual image conversion method based on multi-view feature fusion, effectively transforming event streams into traditional image sequences. Extensive experiments and dataset validations, utilizing multiple evaluation metrics including both global and local assessments, demonstrate the advantages of our proposed method.