Design of a CNN accelerating system for real-time object detection

Abstract

Convolutional neural networks (CNNs) are a type of deep neural networks, that are outstanding in computer vision. Object detection using CNNs has also made great progress in recent years, making it possible to use in real life. These applications often require real-time performance, and dedicated hardware accelerators have been proposed to meet these requirement. However, the proposed CNN accelerators focus on computational efficiency only and do not analyze the overall system performance. Therefore, this thesis proposes a CNN accelerating system that can be easily programmed and achieves real-time performance. In order to easily program the CNN operations, Intuitive basic convolutional primitives are defined and neural processing unit (NPU) is designed. In addition, the control architecture that effectively controls NPU is proposed by analyzing overall system performance from primary input to final output. Through the control optimization, 56% of the processing latency is reduced. The proposed system is implemented on the FPGA development board to verify the object detection for 5 fps.