College - Author 1

College of Engineering

Department - Author 1

Computer Engineering Department

Degree Name - Author 1

BS in Computer Engineering

Date

12-2019

Primary Advisor

John Seng, College of Engineering, Computer Engineering Department

Abstract/Summary

Over the past two years, 3D object detection has been a major area of focus across industry and academia. This is primarily due to the difficulty of learning data from point clouds. While camera images are fixed size and can therefore be easily trained on using convolution, point clouds are unstructured series of points in three dimensions. Therefore, there is no fixed number of features, or a structure to run convolution on. Instead, researchers have developed many ways of attempting to learn from this data, however there is no clear consensus on what is the best method, as each has advantages and disadvantages. For this project, I chose to focus on understanding and implementing VoxelNet, a voxelized method for object detection using point cloud data. I used the VoxelNet architecture for the task of detecting objects in the surrounding environment and creating 3D bounding boxes around those objects. I trained these models on the Waymo Open Dataset, then measured performance on the Carla simulator. The goal of training on the Waymo Open Dataset was to gain experience with the new dataset and familiarity with its features, and then evaluate the practicality of the Carla simulator by using a model trained with real-world data in it.

Included in

Robotics Commons

COinS