Deep neural networks for 2D LIDAR images classification

By

Author

Presented To

Department of Electrical/Electronic Engineering

Abstract 

1 Introduction Theworldhasseenagreatadvanceinthefieldofautonomousvehicles in the past 15 years. In 2003, the well known Tesla, Inc. company— formerly Tesla Motors—was founded in Delaware. In July 2004 The Stanford Racing Team [1] was created under the leadership of professor Sebastian Thrun, who is considered to be the father of self-driving cars. TheteammanagedtowintheDARPAGrandChallengein2005[2] using Stanley, an autonomousVolkswagenTouareg,whichcanbeseen at the Smithsonian’s National Museum of American History in Washington, DC. In 2009 the Google Self-Driving Car Project was founded, lead by Sebastian Thrun, and later became the Waymo company. Both, Tesla and Waymo, specialize in autonomous car developmentbutthey achieve it through different means. Tesla relies on many cameras, radar, and ultrasonics, while Waymo equips its vehicles with 3DVelodyne LiDAR (Light Detection and Ranging) sensor as well as cameras and radar. Both of these approaches require an understanding and fusion of the data from different sensors in real time, which requires a lot of computational power as well as efficient computational models. One of these models are neural networks which are used to segment images from cameras so that the vehicle can comprehend captured camera images and react to them. This is useful, for example, for tracking road surface markings, recognizing traffic lights and traffic signs or detecting other vehicles, pedestrians, etc. In this thesis, I explore the possibility of using neural networks for segmenting 2D LiDAR images for a software company ARTIN and its project Roboauto, which has provided me with recorded LiDAR images.Therearetwosourcesoftheseimages:real-worldenvironment and a simulator. As a part of this thesis, I also process and annotate datafromtherecordings.Icomparedifferentarchitecturesofnetworks trained either on real or simulated data. I also analyze the impact of substituting a part of these networks for a part of an autoencoder which was pretrained on bigger dataset. In Section 2 I present some information about the Roboauto project relevant to this thesis. In Section 3 I introduce the neural networks and also the most common technique for training them -the gradient descent algorithm with backpropagation. In Section 4 I explain the 1