Académique Documents
Professionnel Documents
Culture Documents
Abstract— The blob analysis became a well known method for the detection of objects in
digital images and is an important part in the fields of image processing and computer
vision. Together with the increasing resolutions and frame rates of recent digital video
cameras the analysis requires computationally intensive operations. Software
implementations may not be able to accomplish a satisfying performance. Furthermore,
existing hardware solutions require a processing through the picture in multiple passes. This
paper describes the development of a FPGA algorithm, performing a high speed real-time
blob analysis in only a single pass.
Keywords – blob analysis, object detection, region labelling, image processing, FPGA
n-1
connection between the camera and the FPGA. After
previous row ⇒ 1 1 2 2
current row ⇒ 1 1 a ? n performing the blob analysis, the results are
X X n+1 transferred over the PCI-bus into the main memory
of the host PC. The complete configuration of the
Figure 5: The Merging of two Objects. frame grabber is shown in Figure 6.
The developed algorithm solves this problem in a camera
Silicon SOFTWARE
novel way. As the properties are calculated while frame grabber camera link interface
VII ACKNOWLEDGEMENT
Thanks to Silicon SOFTWARE GmbH [3] which
supports the research as well as provides the frame
grabber hardware and software environment.
VIII REFERENCES
[1] E. Davies, Machine Vision, Academic Press, third
edition, 2005
[2] M.K. Hu, "Visual Pattern Recognition by Moment
Invariants," IEEE Transactions on Information
Figure 8: Simulation Output of an Analysed Image.
Theory, Volume 8, February 1962
Besides the performance, the simulations of [3] Silicon SOFTWARE GmbH, www.silicon-
numerous images have provided an estimation of the software.com
required resources. Here, the crucial factor is the [4] XILINX Spartan II/III, www.xilinx.com, (April 2007)
[5] W. Burger, M.J. Burge, Digitale Bildverarbeitung,
memory required to hold the object properties and Springer Verlag, March 2005.
the memory requirements of the current frame row as [6] R.V. Rachakonda, P.M. Athanas and A.L. Abbott,
well as the previous row. Tests have shown that the "High-Speed Region Detection and Labeling using an
total number of block RAM required is about 72kBit FPGA-based Custom Computing Platform," 5th
depending on the input image and the necessary International Workshop on Field Programmable Logic
object-features. Today's low-cost FPGAs deliver and Applications, Oxford, UK, Sep 1995
between 64k and 280k block RAM bits. Therefore, [7] Camera Link by the Automated Imaging Association,
the algorithm can be easy implemented without any http://www.machinevisiononline.org/public/articles/in
dex.cfm?cat=129, (April 2007)
constraints.
[8] T. Trenschel, "Blob Analyse - Bestimmung von
Formparametern beliebig geformter Objekte auf
VI CONCLUSIONS FPGAs in Echtzeit," Diploma Thesis, Ruprecht-Karls-
Universität, Heidelberg, 2000
This paper has presented the development of a real- [9] ImageJ, http://rsb.info.nih.gov/ij, (April 2007)
time blob analysis algorithm for a FPGA
implementation. The incoming frames from a digital