A mobile cloud framework for deep learning and its application to smart car camera

Abstract

Deep learning has become a powerful technology in image recognition, gaming, information retrieval, and many other areas that need intelligent data processing. However, huge amount of data and complex computations prevent deep learning from being practical in mobile applications. In this paper, we proposed a mobile cloud computing framework for deep learning. The architecture puts the training process and model repository in cloud platforms, and the recognition process and data gathering in mobile devices. The communication is carried out via Git protocol to ensure the success of data transmission in unstable network environments. We used smart car camera that can detect objects in recorded videos during driving as an example application, and implemented the system on NVIDIA Jetson TK1. Experimental results show that detection rate can achieve four frame-per-second with Faster R-CNN and ZF model, and the system can work well even when the network connection is unstable. We also compared the performance of system with and without GPU, and found that GPU still plays a critical role in the recognition side for deep learning.