Novel memory systems tailored to vision algorithms

Abstract

Nowadays, one of the most important challenges in the field of Computer Vision and the target of all the engineers that are associated with this, is the design of embedded systems that will execute real-time algorithms with big accuracy and low cost. This effort meets a lot of difficulties as the algorithms that have already been applied “suffer” by bottlenecks, especially as far as that concerns the memory accesses. In this thesis, we deal with a novel embedded system, which implements the OpenTLD algorithm and can be applied at a broad-range of algorithms based on the generalized Viola and Jones framework. Viola and Jones framework, despite the fact that is hardware friendly in resource-limited devices (FPGAs), the way it access the memory causes certain problems and parallelism is not easy to be applied on it. This is mainly due to the fact that it is a memory bound problem and its memory access pattern has certain characteristics that make it hard to take advantage of the conventional memory hierarchies. Our system accelerates the bottleneck of the algorithm with a high bandwidth distributed memory sub-system which is independent of the various software parameters. Thus, it was of high interest, the effort to try this embedded system on an algorithm out of Viola and Jones framework and study the result. That is the second part of the thesis. The first one is to optimize the first edition of the system in order to be more efficient and easier as concerns the communication with different algorithms.