Analyzing 2D-Convolutional Neural Network's Ability to Learn Obstacle Avoidance in Self-Driving Applications
CSEF · 2018 Mathematical Sciences
Overview
Objectives/Goals Analyze the ability of a 2D Convolutional Neural Network to learn random 3D objects as obstacles by training it with fixed set of 3D objects of varying color and orientation placed randomly on the track. Methods/Materials Added a Hall Effect sensor and related code changes to the "Donkey" open source based self-driving car project which lacked consistent speed control and precise maneuverability so that it could be used with my project. After mastering repeatable autonomous driving around a test track loop, several iterations of training placing a pre-selected list of various 3D objects and avoiding them was carried out. Training included systematically varying the independent variables such as color of the given object, its orientation and its position on the track. After training, a random set of new 3D objects where used to evaluate the learning achieved in general and to see how each independent variable impacted the effectiveness of learning in a quantitative manner. Experiments were carried out in controlled environment (light condition, speed, track etc.) to ensure the results observed could be repeated. Results When 3D objects used during training and random new 3D objects not seen before by the car were randomly placed at different orientations and at random places on the track during testing, the car avoided them as obstacles in most conditions, failed to maneuver effectively or failed to detect the object as obstacles in others. Conclusions/Discussion The experiment clearly demonstrated that 2D CNN with a single camera input could be used to achieve obstacle avoidance. However, the effectiveness of the results depended upon specific aspects of the independent variables. For example, darker objects were avoided much predictably than lighter colored ones. Obstacles were avoided consistently on straight sections than on curves. Both quality and quantity of training had a great impact on the observed behavior of a run.
Summary statement
The obstacle avoidance behavior observed during testing with random new 3D objects which differed significantly from the training objects clearly demonstrated 2D CNN's ability to learn obstacle avoidance.
Help received
I built the car per Donkey project and added Hall Sensor. Discussed new concepts with father. My mentor helped point to useful research. Parents helped placing objects on track as I trained.
Competition history
- CSEF 2018
Resources
Related projects
CSEF · 2018
Deep Learning Based Collision Avoidance Algorithm for Mobile Robots in Pedestrian Environments
CSEF · 2019
Obstacle Avoiding Robot: A Reinforcement Learning Approach
ISEF · 2018
Using Deep Convolutional Neural Networks and Feature Matching Algorithms to Develop Smarter Autonomous Cars
ISEF · 2018
CARL: A Convolutional Neural Network Powered Self-Driving Car
CSEF · 2010
Utilizing Artificial Neural Networks to Create a Learning Robot
CSEF · 2018
Deep Learning Real-Time Object Detection through Convolutional Neural Networks Using OpenCV for the Visually Impaired
CSEF · 2012
Autonomous Navigation and Robust Object Detection through Edge-based Optical Flow
CSEF · 2016
Improved Machine Perception During Autonomous Driving via Semisupervised Deep Learning and Multisensor Datastream Fusion
Closest projects by meaning, across every fair and year in the corpus.
Browse more like this
Source: California Science & Engineering Fair public projects