Exploiting and Benchmarking Hardware Acceleration for Computer Vision on Industrial Embedded Devices

In this thesis, readily available commercial solutions for AI at the edge are discussed, and two Systems on a Chip (SoC) with integrated AI accelerators from two distinct producers are assessed. Capabilities, performance, flexibility and the systems as a whole will be evaluated and depicted in the broader context of state-of-the-art innovations. Furthermore, an analysis will be conducted on a tensor-oriented algorithm needed as a post-processing step for a particular AI model. Optimizations will be conducted to improve performance on lower-end CPUs present on embedded devices, whose features include standard and widespread ISA extensions like Arm® Neon™ so that a speedup can be obtained by taking advantage of vector and SIMD instructions. This thesis work is the result of a four-month internship experience at Deep Vision Consulting (www.deepvisionconsulting.com).