Road Scene Understanding Presented by: Mohamed Mohsen Agenda - PowerPoint PPT Presentation

Road Scene Understanding Presented by: Mohamed Mohsen

Agenda ● Problem Definition ● Currently existing solutions ● Thesis statement ● Questions to answer ● Introduction about deep learning

Problem Definition ● Detect roads with and without lane markings using cameras ● Needed for autonomous driving ● Current solutions are extremely expensive

Currently used solutions 64 Laser Velodyne LIDAR System

Currently used solutions 64 Laser Velodyne LIDAR System

Currently used solutions ● Camera Based solutions have not reached satisfiable accuracy ● Best accuracy achieved (using the KITTI dataset) is 91.76%

KITTI Dataset Kitti Car

Thesis Statement Improve the accuracy of road detection using cameras

Questions to answer ● What is the most suitable technique to use? – Proposal: Deep learning ● What is the best model to use for feature representation? ● What is the best network architecture for this problem?

Introduction to Deep Learning ● Developed to address problems people found with other techniques like backpropagation ● Addresses the problem of data representation – Computers are better in finding the representations than humans ● Emulates the way our brain works by having consecutive layers each one computing a higher level function based on the output of the previous layer

Introduction to Deep Learning Traditional Back Propagation

Introduction to Deep Learning Representation Learning

Introduction to Deep Learning Deep Learning Model

T h a n k Y o u