Date of Award


Publication Type

Doctoral Thesis

Degree Name



Electrical and Computer Engineering




Q. M. J. Wu




Today, the implementation of machine vision algorithms on embedded platforms or in portable systems is growing rapidly due to the demand for machine vision in daily human life. Among the applications of machine vision, human action and activity recognition has become an active research area, and market demand for providing integrated smart security systems is growing rapidly. Among the available approaches, embedded vision is in the top tier; however, current embedded platforms may not be able to fully exploit the potential performance of machine vision algorithms, especially in terms of low power consumption. Complex algorithms can impose immense computation and communication demands, especially action recognition algorithms, which require various stages of preprocessing, processing and machine learning blocks that need to operate concurrently. The market demands embedded platforms that operate with a power consumption of only a few watts. Attempts have been mad to improve the performance of traditional embedded approaches by adding more powerful processors; this solution may solve the computation problem but increases the power consumption. System-on-a-chip eld-programmable gate arrays (SoC-FPGAs) have emerged as a major architecture approach for improving power eciency while increasing computational performance. In a SoC-FPGA, an embedded processor and an FPGA serving as an accelerator are fabricated in the same die to simultaneously improve power consumption and performance. Still, current SoC-FPGA-based vision implementations either shy away from supporting complex and adaptive vision algorithms or operate at very limited resolutions due to the immense communication and computation demands. The aim of this research is to develop a SoC-based hardware acceleration workflow for the realization of advanced vision algorithms. Hardware acceleration can improve performance for highly complex mathematical calculations or repeated functions. The performance of a SoC system can thus be improved by using hardware acceleration method to accelerate the element that incurs the highest performance overhead. The outcome of this research could be used for the implementation of various vision algorithms, such as face recognition, object detection or object tracking, on embedded platforms. The contributions of SoC-based hardware acceleration for hardware-software codesign platforms include the following: (1) development of frameworks for complex human action recognition in both 2D and 3D; (2) realization of a framework with four main implemented IPs, namely, foreground and background subtraction (foreground probability), human detection, 2D/3D point-of-interest detection and feature extraction, and OS-ELM as a machine learning algorithm for action identication; (3) use of an FPGA-based hardware acceleration method to resolve system bottlenecks and improve system performance; and (4) measurement and analysis of system specications, such as the acceleration factor, power consumption, and resource utilization. Experimental results show that the proposed SoC-based hardware acceleration approach provides better performance in terms of the acceleration factor, resource utilization and power consumption among all recent works. In addition, a comparison of the accuracy of the framework that runs on the proposed embedded platform (SoCFPGA) with the accuracy of other PC-based frameworks shows that the proposed approach outperforms most other approaches.