Term of Award

Summer 2018

Degree Name

Master of Science in Applied Engineering (M.S.A.E.)

Document Type and Release Option

Thesis (restricted to Georgia Southern)

Copyright Statement / License for Reuse

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Department

Department of Mechanical Engineering

Committee Chair

Biswanath Samanta

Committee Member 1

Brian Vlcek

Committee Member 2

JungHun Choi

Abstract

A study is presented on the use of deep neural network (DNN) systems for object detection and distance estimation in autonomous robotic navigation. A mobile robot, Turtlebot, outfitted with a fast, power-efficient embedded DNN computing Nvidia Jetson TX2 processor was used. A single red-green-blue-depth (RGB-D) camera was used to evaluate three methods of estimating the distance of objects and obstacles within the framework of Robot Operating System (ROS). The depth information of the RGB-D camera and two DNN based algorithms, namely, MonoDepth and Tensorflow API’s single-shot detector (SSD) MobileNet with RGB images were used for distance estimation. The SSD MobileNet system was tested through people while the RGB-D and MonoDepth system were tested through both people and black boards as obstacles/objects. The effects of camera angle, elevation and person’s height on accuracies of distance estimation were studied. The effectiveness of distance estimation processes was demonstrated in autonomous navigation of the robotic platform with obstacle avoidance (brake test) and simple path planning. Static distance estimation results of RGB-D were the most consistent and accurate in the effective range of 0-13m. Static distance estimations of both MonoDepth and Tensorflow DNNs were within 10% of error for distance range of 0-35m 90% of the time. During the navigation tests using RGB-D method, the additional lag, possibly due to ROS, made the robot move an extra distance in brake tests, though within 10% error. Otherwise, the performance of RGB-D was most consistent in path planning. The performance of MonoDepth DNN was relatively inconsistent in brake tests and path planning, though mostly within allowable error. The performance of Tensorflow API SSD MobileNet in brake tests and path planning was with little lag and within allowable error. Overall, all three methods were found to work within allowable error limit for autonomous robotic navigation though DNN based Tensorflow API SSD MobileNet could be considered as a potential option for robotic perception.

Research Data and Supplementary Material

No

Available for download on Saturday, August 31, 2019

Share

COinS