Stereo matching algorithms are used for finding relative depth from a pair of rectified stereo images. The resultant disparity information can be used for 3D reconstruction by triangulation, using the known intrinsic and extrinsic parameters of the stereo camera. The Semi global method for stereo disparity estimation aggregates the cost in terms of dissimilarity across multiple paths leading to a smoother estimate of the disparity map.
For the semi-global method in Vitis Vision, census transform in conjunction with Hamming distance is used for cost computation. The semiglobal optimization block is based on the implementation by Hirschmuller, but approximates the cost aggregation by considering only four directions.
Parallelism is achieved by computing and aggregating cost for multiple disparities in parallel, and this parameter is included as a compile-time input.
API Syntax
template<int BORDER_TYPE, int WINDOW_SIZE, int NDISP, int PU, int R, int SRC_T, int DST_T, int ROWS, int COLS, int NPC, int XFCVDEPTH_IN_L = _XFCVDEPTH_DEFAULT, int XFCVDEPTH_IN_R = _XFCVDEPTH_DEFAULT, int XFCVDEPTH_OUT = _XFCVDEPTH_DEFAULT>
void SemiGlobalBM(xf::cv::Mat<SRC_T,ROWS,COLS,NPC,XFCVDEPTH_IN_L> & _src_mat_l, xf::cv::Mat<SRC_T,ROWS,COLS,NPC,XFCVDEPTH_IN_R> & _src_mat_r, xf::cv::Mat<DST_T,ROWS,COLS,NPC,XFCVDEPTH_OUT> & _dst_mat, uint8_t p1, uint8_t p2)
Parameter Descriptions
The following table describes the template and the function parameters.
| Parameter | Description |
|---|---|
| BORDER_TYPE | The border pixels are processed in Census transform function based on this parameter. Only XF_BORDER_CONSTANT is supported. |
| WINDOW_SIZE | Size of the window used for Census transform computation. Only ‘5’ (5x5) is supported. |
| NDISP | Number of disparities |
| PU | Number of disparity units to be computed in parallel |
| R | Number of directions for cost aggregation. It must be 2, 3, or 4. |
| SRC_T | Type of input image Mat object. It must be XF_8UC1. |
| DST_T | Type of output disparity image Mat object. It must be XF_8UC1. |
| ROWS | Maximum height of the input image. |
| COLS | Maximum width of the input image. |
| NPC | Number of pixels to be computed in parallel. It must be XF_NPPC1. |
| XFCVDEPTH_IN_L | Depth of the input image. |
| XFCVDEPTH_IN_R | Depth of the input image. |
| XFCVDEPTH_OUT | Depth of the output image. |
| _src_mat_l | Left input image Mat |
| _src_mat_r | Right input image Mat |
| _dst_mat | Output disparity image Mat |
| p1 | Small penalty for cost aggregation |
| p2 | Large penalty for cost aggregation. The maximum value is 100. |
Resource Utilization
The following table summarizes the resource utilization for a 1920 x 1080 image, with 64 number of disparities, and 32 parallel units.
| Operating Mode | Filter Size | Operating Frequency (MHz) | Utilization Estimate | |||
|---|---|---|---|---|---|---|
| BRAM_18K | DSP_48Es | FF | LUT | |||
| 1 Pixel | 5x5 | 200 | 205 | 141 | 11856 | 19102 |
Performance Estimate
The following table summarizes a performance estimate for a 1920x1080 image.
| Operating Mode | Operating Frequency | Number of Disparities | Parallel Units | Latency |
|---|---|---|---|---|
| 1 pixel/clock | 200 MHz | 64 | 32 | 42 ms |