Accelerated Application – Open VSLAM

Supported by AMD for KR260, the algorithm of Open VSLAM (Vision based Simultaneous Localization And Mapping) offers the possibility of purely vision-based 3D mapping.


With the help of cameras, in our case two, the own location on a map is recognized, a 3D map is created, and the circle is closed by loop closure. Tools like GPS, gyro or IMU are neither needed for map creation nor for depth image.

The most computationally
intensive steps are

The processing steps of OpenVSLAM are displayed schematically in the graphic (OpenVSLAM in short blocks). All steps from the input image to the map generation including loop detection and localization are shown.


Depth estimation (with color space conversion, rectification and the actual matching)


Keypoint detection (here an ORB – oriented FAST and rotated BRIEF) based on

—   BRIEF (Binary Robust Independent Elementary Features): Rotation invariant
—   FAST (Features from Accelerated Segment Test): Orientated in terms of scale

Open VSLAM in short blocks


Bundle Adjustments**

ROS graph, based on initial
steps of FPGA acceleration

Euroc is the input-frame, which we take before the keypoint detection to calculate the depth image, which will be used later in triangulation of 3D points. (See also Resources / Downloads: Paper Open VSLAM in Short Blocks)