Evaluating Resources Cost of a Convolutional Neural Network Aiming an Embedded System

摘要

The use of Machine Learning algorithms in image classification problems have yielded satisfactory results in recent years. Classification algorithms such as Support Vector Machines (SVMs) combined with robust feature extractors like Histogram of Oriented Gradients (HOG) have been used to achieve accuracy results over 95%. Very recently, with the researches applied in the deep learning fields, Convolutional Neural Networks (CNNs) have shown to work extremely well with data that has high dimensionality like images. This work focuses on evaluating the resources costs of deploying a CNN in an embedded platform to solve the people detection problem. An implementation of the CNN classification algorithm was developed, and tests were carried out both in a PC and in an embedded platform. Furthermore, a study on the amount of memory and time spent by a classic CNN was executed. The results point out that new networks must be designed to fit in low-resources embedded platforms.