"Lightweight Convolutional Neural Networks for Camera-Based Object Detection on Mobile Devices"
Lightweight Convolutional Neural Networks for Camera-Based Object Detection on Mobile Devices
Sprache des Titels:
Nowadays, convolutional neural networks (CNNs) are widely used in various applications. Especially, tasks like speech and text recognition as well as object detection are very well suited for these kind of networks. The latter can be divided into two subtasks, namely finding objects in an image and determine their location. Of course, there are classical approaches, like pattern matching, to solve these tasks. Unfortunately, these approaches are often prone to error due to scaling and rotation of the objects to be detected. Furthermore, it is time consuming and computationally expensive to implement pattern matching for several different objects. On the other hand, CNNs are well suited for this task due to their generality and applicability for lots of different objects. Nevertheless, due to their high computational costs, these networks are typically operated on workstations with fast central processing units (CPUs) and, especially graphic cards. Anyhow, it is possible to train a CNN on a machine providing high computational power and utilize the trained network on mobile devices to perform the actual object detection. In this thesis the performance in terms of speed of lightweight CNNs for camera based object detection on mobile devices is investigated. For this purpose, three of the most popular state of the art lightweight object detection networks are examined. In addition, a Python script as well as an Android application capable of executing pretrained neural networks are presented. Based on this implementation the CNNs are compared in terms of speed measured in frames per second (FPS).