Researchers have developed a binocular visual SLAM system based on event cameras that addresses the limitations of existing methods, which often rely on long event stream accumulation and fail to exploit correlations between events, leading to poor mapping and localization accuracy. The proposed system matches events with the nearest timestamps on left and right time surfaces to compute feature depth, then uses the Lucas-Kanade algorithm to estimate small pose increments for tracking. Depth information is fused using an iterative reweighted least squares (IRLS) algorithm based on a Student's t probability model, producing a semi-dense map.
Key takeaways
- Pose error reduced by 48.3% compared with state-of-the-art methods across different datasets.
- Robust performance in low-light and high-dynamic scenes, where traditional visual SLAM often fails.
- IRLS-based fusion outperforms least squares in mapping density and accuracy, with lower error variance.
Tests on open-source datasets simulating extreme environments demonstrated superior mapping quality and higher localization precision. The approach leverages the asynchronous nature of event cameras, using time surfaces and spatiotemporal-aware event processing to avoid filtering out true events. This work offers a practical path for deploying event-camera SLAM in industrial automation, particularly in challenging lighting or high-speed scenarios where conventional cameras struggle.
Source: 《机器人》期刊 (robot.sia.cn) · Published 2026-01-13 · “基于事件相机的双目视觉SLAM”