Semi-Dense Scene Reconstruction Based on Stereo Event Cameras

Compared to standard cameras, event cameras that work asynchronously offer many advantages, such as low power consumption, high temporal resolution, and high dynamic range, making them valuable tools for tasks like real-time 3D reconstruction, simultaneous localization and mapping (SLAM), and more. However, due to the asynchronous event output characteristic and their photometric sensing method, traditional 3D reconstruction algorithms are challenging to apply directly. Existing reconstruction methods based on stereo event cameras often face issues such as pose estimation drift and insufficient reconstruction accuracy in complex motion scenarios. To address this, this paper builds a semi-dense 3D reconstruction system based on stereo event cameras, comprehensively compares four commonly used event stream processing methods, and proposes an innovative event representation method Adaptive Time Surface. This method uses an exponential decay kernel to convert event streams into distance field-like event image frames, allowing adaptive parameter adjustment to accommodate differ-ent event frequencies, thereby generating event representations with stable brightness. This significantly enhances the robustness of pose estimation and the accuracy of scene reconstruction. Extensive experiments in real-world scenarios validate that the 3D reconstruction system based on ATS demonstrates stronger robustness and higher accuracy in camera pose estimation and scene reconstruction compared to other event stream processing methods.