Real-time invariant textural object recognition with FPGAs

An unsolved problem in machine vision is the real time extraction of object type and position in a noisy three dimensional environment without additional non visual information. Existing algorithms do not scale well to large model databases, as would be required for true three dimensional recognition in the real world. This paper proposes a new algorithm and method of implementation to solve this problem. The new system is based on four Spartan 3A FPGAs, and is capable of real time textural object recognition at 320times240 pixels in a noisy environment. In addition, the system contains rudimentary depth perception when provided with more than one camera, and is able to work with and store a large model database in nonvolatile memory. The base algorithm, hardware design, and testing schemes will be discussed in detail. Finally, it will be shown that this system is superior to prior art in this field.