Efficiently mapping high-performance early vision algorithms onto multicore embedded platforms

• Main Author: Apewokin, Senyo
• Published: Georgia Institute of Technology 2009
• Subjects: Computer vision; Embedded; Multicore; Algorithms
• Online Access: http://hdl.handle.net/1853/28256

The combination of low-cost imaging chips and high-performance, multicore, embedded processors heralds a new era in portable vision systems. Early vision algorithms have the potential for highly data-parallel, integer execution. However, an implementation must operate within the constraints of embedded systems including low clock rate, low-power operation and with limited memory. This dissertation explores new approaches to adapt novel pixel-based vision algorithms for tomorrow's multicore embedded processors. It presents : - An adaptive, multimodal background modeling technique called Multimodal Mean that achieves high accuracy and frame rate performance with limited memory and a slow-clock, energy-efficient, integer processing core. - A new workload partitioning technique to optimize the execution of early vision algorithms on multi-core systems. - A novel data transfer technique called cat-tail dma that provides globally-ordered, non-blocking data transfers on a multicore system. By using efficient data representations, Multimodal Mean provides comparable accuracy to the widely used Mixture of Gaussians (MoG) multimodal method. However, it achieves a 6.2x improvement in performance while using 18% less storage than MoG while executing on a representative embedded platform. When this algorithm is adapted to a multicore execution environment, the new workload partitioning technique demonstrates an improvement in execution times of 25% with only a 125 ms system reaction time. It also reduced the overall number of data transfers by 50%. Finally, the cat-tail buffering technique reduces the data-transfer latency between execution cores and main memory by 32.8% over the baseline technique when executing Multimodal Mean. This technique concurrently performs data transfers with code execution on individual cores, while maintaining global ordering through low-overhead scheduling to prevent collisions.