Investigation on performance and energy efficiency of CNN-based object detection on embedded device

The use of a Convolutional Neural Network based method for object detection increases the accuracy that surpasses human visual system. Because it requires considerable computational capability, its use in embedded devices that place constraints in terms of power consumption as well as computational capability has thus far been limited. However, with the recent development of GPU for use in embedded devices and open-source software library for machine learning, it has become viable to utilize CNN in an energy-efficient embedded computing environment. In this study, CPU and GPU performance and energy efficiency of CNN-based object detection inference on an embedded platform is investigated through comparison with a traditional PC-based platform. Two publicly available hardware platforms are empirically evaluated; in one of them-NVIDIA Jetson TX-1-the results demonstrate image processing performance of 65% of that of the PC, while the embedded device consumes 2.6% of power consumed by the PC.