The stereo vision IP core for FPGAs is the heart of our SceneScan stereo vision sensor. We are happy to offer you this IP core for licensing for integration of our stereo vision capabilities into your own FPGA-based products.
To facilitate the use of the stereo vision core on devices with shared system memory, we also offer an IP core for Direct Memory Access (DMA). This DMA core reads the input data required by the stereo vision core from the system memory via AXI3 and converts them into a data stream suitable for the stereo vision core. The DMA core also collects the processing results and writes them back into the system memory. The DMA core is particularly suitable for use on the Xilinx Zynq SoC.
The stereo vision IP core performs stereo matching on two monochrome or RGB input images. The images are first rectified inside the FPGA to compensate for lens distortions and camera alignment errors. Stereo matching is then performed by applying a variation of the Semi Global Matching (SGM) algorithm, which has been optimized for FPGA-based processing. Various post-processing methods are applied to improve the computed depth data. The output of the stereo vision core is a subpixel accurate and dense disparity map (an inverse depth map), which is streamed over an AXI4-Stream interface.
- Processing of grayscale images with a bit depth of 8 or 12 bits per pixel
- Processing of color images with 8-bit RGB or Bayer pattern encoding
- Stream-based processing of input images using either AXI4-Stream, AXI4 or AXI3
- Configuration through AXI4-Lite interface
- Output of disparity map starts before receiving the last pixel of both input images
- Support for variable image sizes
- Multi-clock design with faster clock for performance critical tasks
- Rectification using a pre-computed compressed rectification map
- Bi-linear interpolation for subpixel accurate rectification
- Stereo matching through a variation of the Semi-Global Matching (SGM) algorithm
- Configurable disparity range up to 256 pixels
- Configurable disparity offset
- Configurable penalties P1 and P2 for small and large disparity variations
- Edge-dependent variation of penalties
- Pre-processing of input images for improved robustness against illumination variations and occlusions
- Subpixel optimization
- Consistency check with configurable threshold
- Uniqueness check with configurable threshold
- Filling of small gaps through interpolation
- Noise reduction
- Speckle filtering
- Filtering of untextured image areas