Smart Camera using FPGA

Abstract

As we experience the third wave of artificial intelligence, which entails contextual adaptation, technology is advancing with implementation of machine learning algorithms in many applications. In such applications, deep learning, particularly Convolutional Neural Networks have proved to be highly effective in objective detection, classification and identification. A landmark achievement was the high accuracy of CNNs in the ImageNet competition. 2 Training these networks, although computation-intensive, does not require real time processing and can be carried out using traditional processing systems or even GPUs. However, inference of these algorithms requires real-time computations, which not only reduce power consumption but also optimize the implementation of CNNs via parallelism and optimum memory placement. FPGA-based architectures are the most optimum, as of yet, for implementation of CNN algorithms. To prove this, we aim to implement a deep learning algorithm on Zynq-7000 to prove the effectiveness of FPGA portability, configurability and power efficiency as opposed to other architectures.