Candidate
Zhizhou Li
Title
Study of Implementation of CNN on Low-power Platform for Smart Traffic Optimization
Supervisors
Kshirasagar Naik and Justin Eichel (Systems Design)
Abstract
Accompanied with the rise of smart city and the development of IoT (Internet of Things), people are looking forward to monitoring and regulating the traffic in a smarter way. Since the deep neural network has shown its great value in vehicle detection area, people may wonder what kind of impact would brought by the combination of IoT and deep learning techniques. In this work, an exploration of implementation of CNN (convolutional neural network) on low-power platform for smart traffic optimization has been launched. During the research, a new optimization approach, which aims at S-CNN (Sparse Convolutional Neural Network) optimization from architecture level, has been proposed and obtain outstanding performance when compared to mainstream deep learning framework, such as Tensorflow. In the experiments, the new proposed S-CNN optimization approach is as twice fast as Tensorflow on 94% sparse model and becomes 5 times faster on 98% sparse model. Besides, the author also verified the feasibility of real-time CNN implementation on ARM platform and Jetson TX1 embedded system, which reveals the shortage of computational resource on ARM platform and the potential of Jetson series to become the low-power platform for CNN implementation.