intelligent cameras and embedded reconfigurable computing a.pptVIP

Binarization (again) and Merging The resulting image is then binarized again, since the aim of the application is not to detect the magnitude of the gradient but the presence of a gradient. Finally, the edge detected image is merged with the original image. For that goal, inverse binarization is applied i.e., the background is filled by 1’s and moving image edges by 0’s, thus allowing to implement the merge operation with a multiplication. Software-Only performance evaluation Motion Detection on DREAM The implementation, or in general the acceleration (of the application under consideration) on the reconfigurable platform is driven by two main factors: Instruction/Data level Parallelism: Each operation (e.g. binarization, edge detection..) show relevant instruction level parallelism. Given a specific image processing kernel, computation associated to each pixel is independent from the elaboration of other pixels, although the reuse of adjacent pixels prove beneficial to minimize memory access. Data Size: After binarization, the information content associated to each pixel can be represented by only 1 bit (edge/No-edge), thus allowing to store upto 32 pixels in a 32 bit word. This significantly reduces memory utilization without implying additional packing/unpacking overhead, as it would be the case with 32-bit processors, since DREAM may handle shifts by programmable routing. Benefits of reduced “data size” The last consideration (reduced data size) provides additional benefits since: The erosion phase requires the search of the minimum within the pixels in a 3x3 matrix, but is implemented on DREAM by a single 9-bit input 1-bit output AND; The dilation phase requires the search of the maximum within the pixels in a 3x3 matrix, but is implemented on DREAM by a single 9-bit 1-input OR; Edge detection can be strongly simplified as explained in the following slides: Inverse-Binarized edge detection Inverse binarized edge detection can be represented by: Since, each