Energy efficient enabling technologies for semantic video processing on mobile devices
Larkin, Daniel (2008) Energy efficient enabling technologies for semantic video processing on mobile devices. PhD thesis, Dublin City University.
Full text available as:
Semantic object-based processing will play an increasingly important role in future multimedia systems due to the ubiquity of digital multimedia capture/playback technologies and increasing storage capacity. Although the object based paradigm has many undeniable benefits, numerous technical challenges remain before the applications becomes pervasive, particularly on computational constrained mobile devices. A fundamental issue is the ill-posed problem of semantic object segmentation. Furthermore, on battery powered mobile computing devices, the additional algorithmic complexity of semantic object based processing compared to conventional video processing is highly undesirable both from a real-time operation and battery life perspective. This
thesis attempts to tackle these issues by firstly constraining the solution space and focusing on the
human face as a primary semantic concept of use to users of mobile devices. A novel face detection algorithm is proposed, which from the outset was designed to be amenable to be offloaded from the host microprocessor to dedicated hardware, thereby providing real-time performance and
reducing power consumption. The algorithm uses an Artificial Neural Network (ANN), whose topology and weights are evolved via a genetic algorithm (GA). The computational burden of the ANN evaluation is offloaded to a dedicated hardware accelerator, which is capable of processing
any evolved network topology. Efficient arithmetic circuitry, which leverages modified Booth recoding, column compressors and carry save adders, is adopted throughout the design. To tackle the increased computational costs associated with object tracking or object based shape encoding, a novel energy efficient binary motion estimation architecture is proposed. Energy is reduced in the proposed motion estimation architecture by minimising the redundant operations inherent in the binary data. Both architectures are shown to compare favourable with the relevant prior art.
Archive Staff Only: edit this record