US20260079009
2026-03-19
Physics
G01C21/206
The patent application discusses a method for simultaneous localization and mapping (SLAM) using dual event-based cameras. These cameras generate event streams that are processed to detect surface points in an environment, compute the camera's pose dynamically, and update the map concurrently. The approach uses gradient descent optimization to refine the pose for each event or small batches of events. This technique is part of the field of computer vision and aims to enhance SLAM by leveraging the capabilities of event cameras.
SLAM is a process used by devices, such as robots or drones, to map an environment while determining their position relative to it. Traditional SLAM systems often face challenges like scale ambiguity and drift, particularly in monocular setups that rely on single cameras. These systems typically require additional data sources for depth measurement and suffer from delayed feature initialization. The dual event camera approach seeks to address these limitations by providing more reliable scale information and reducing the need for special initialization phases.
The dual event cameras are positioned to have overlapping fields of view, which enables stereoscopic detection of surface points and depth measurement via epipolar geometry. The image processing system utilizes these overlapping views to compute the depths of features and dynamically update the camera's pose and the environment's map. By using event-based data, the system can efficiently handle high-speed motion and reduce latency and power consumption compared to traditional video-based SLAM systems.
This method offers several advantages over conventional SLAM systems, including reduced power consumption, latency, and jitter, as well as increased robustness to high-speed motion. The approach is beneficial in various applications such as virtual reality, augmented reality, automotive navigation, drone navigation, and domestic robotics. By simultaneously estimating the camera's pose and mapping the environment, it enhances the robustness and accuracy of SLAM systems in dynamic environments.
The system is implemented with two event cameras, each providing asynchronous event streams that are processed to generate depth images. The cameras are calibrated with known separation and orientation, allowing the image processing system to calculate distances to objects by analyzing the relative positions of event frames. This setup enables the creation of a 3D map and accurate pose estimation, crucial for navigation and interaction with the environment.