Title

Teaching an Autonomous Vehicle to Detect and Report Defects on Road Surface

Document Type

Student Presentation

Presentation Date

4-15-2019

College

College of Engineering

Department

Department of Mechanical & Biomedical Engineering

Faculty Sponsor

Dr. Yang Lu

Abstract

The goal of this project is to create an autonomous pavement surface crack inspection platform that combines autonomous driving and intelligent distress detection. Our small-scale autonomous vehicle can maneuver its way while collecting images of pavement condition for crack detection. It incorporates an open-source library named 'Donkey Car' which runs on TensorFlow for the auto driving control system. It allows users to train their own vehicles according to specific road condition and ambient environments. We built a test track in the Smart Infrastructure Lab and started training the autopilots architecture. We drove it around the track many times with the help of first-person view using local web server control environment. It captures images and corresponding parameters like throttle and orientation to adapt to the roadway conditions, e.g. straight lane, turning points, and uneven conditions. After training, the vehicle can reasonably mimic a safe driving behavior to collect condition data around our test track.

After a successful collection of condition data, we built a deep learning-enabled model to identify and report pavement surface cracks. To implement real-time crack detection, we added a Nvidia Jetson TX2 neural network processor on the vehicle. Then we created a video stream pipeline to capture the high-resolution images from TX2's on-board camera. Subsequently, we annotated over 600 pavement condition images to create training database for the model. Meanwhile, we have set up a data training environment for object detection in TensorFlow based on Single Shot Multibox Detector Architecture on the R2 High Performance Cluster to accelerate the image data training process by the GPU-enabled computing system. The preliminary testing data demonstrated the feasibility this proposed work where inspection will be autonomous and less costly. The overall crack detection accuracy reached about 91%. Currently, we are working on creating a model that can classify different types of cracks on pavement surface. This project has many broader application impacts in the transportation system monitoring and management, including highway system maintenance management, connected vehicle autonomous control, safe and comfortable driving behavior, and cybersecurity testbed for cyber-physical systems.

This document is currently not available here.

Share

COinS