Accelerating Design and Implementation of Embedded Vision Systems

Open Access
Al Maashri, Ahmed
Graduate Program:
Computer Science and Engineering
Doctor of Philosophy
Document Type:
Date of Defense:
September 27, 2012
Committee Members:
  • Vijaykrishnan Narayanan, Dissertation Advisor
  • Yuan Xie, Committee Member
  • Robert Collins, Committee Member
  • Richard A Behr, Special Member
  • Embedded Vision Systems
  • Domain-Specific Computing
  • Reconfigurable Computing
  • Power Efficiency
  • Neuromophic Accelerators
In today’s world, embedded systems have become a necessity in our daily lives. Ranging from digital watches to factory controllers, these systems are dedicated to handle a particular task both efficiently and reliably. The advancements in computer vision and machine learning, in conjunction with the abundance of computational power, made it possible for these systems to perform other tasks such as image processing and video analytics; giving rise to embedded vision systems. Being the heart of embedded vision systems, computer vision is the key technology enabler for image processing and video analytics. However, the computational modalities that are prominent in computer vision algorithms are usually inconsistent with those modalities most exploited in contemporary computer architectures. As a result, execution inefficiencies are observed when algorithms are processed by such architectures. Furthermore, recent studies have shown that biologically-inspired (neuromorphic) vision algorithms can be robust alternatives, due to their detection and recognition capabilities. Interestingly, these neuromorphic algorithms experience similar execution inefficiencies when running on general purpose processors. On the other hand, domain-specific computing is believed to be the solution to the challenges presented above. Developing customized hardware accelerators targeting specific workloads, is the key to achieving the desired performance while operating at a lower power budget. However, domain-specific computing may be an unfavorable route to many, due to a whole host of challenges associated with hardware design and implementation. This dissertation addresses the issues presented above by describing a hardware-software framework that offers a flexible, reliable, and high performance accelerator infrastructure. The underlying hardware complexity of the framework is abstracted from the user through a standardized Application Programming Interface, API. Furthermore, the dissertation presents a software automation tool for expediting the process of building embedded vision systems and mapping them to prototyping platforms. In addition, this dissertation discusses the hardware architecture of several vision accelerators, including neuromorphic accelerators, which are mapped to the accelerator framework described above. In particular, this dissertation presents accelerators for a neuromorphic vision algorithm, HMAX. This algorithm can be used as a feature extractor for multiple recognition tasks. Results reveal that the neuromorphic accelerators can deliver as much as 7.6X speedup, and are up to 12.8X more power efficient when compared to contemporary CPU and GPU platforms. Additionally, results indicate that the neuromorphic accelerators can achieve up to 90% classification accuracy on some of the standard datasets. Furthermore, this dissertation discusses the hardware architecture of other non-neuromorphic embedded vision systems, where results show that these accelerators can speed up the execution time of some of the computer vision algorithms by up to 100X when compared to a CPU platform.