Document Type

Thesis - Open Access

Award Date

2019

Degree Name

Master of Science (MS)

Department

Electrical Engineering and Computer Science

First Advisor

Kwanghee Won

Abstract

The instance segmentation and object detection are important tasks in smart car applications. Recently, a variety of neural network-based approaches have been proposed. One of the challenges is that there are various scales of objects in a scene, and it requires the neural network to have a large receptive field to deal with the scale variations. In other words, the neural network must have deep architectures which slow down computation. In smart car applications, the accuracy of detection and segmentation of vehicle and pedestrian is hugely critical. Besides, 2D images do not have distance information but enough visual appearance. On the other hand, 3D point clouds have strong evidence of existence of objects. The fusion of 2D images and 3D point clouds can provide more information to seek out objects in a scene. This paper proposes a series of fronto-parallel virtual planes and inverse perspective mapping of an input image to the planes, to deal with scale variations. I use 3D point clouds obtained from LiDAR sensor and 2D images obtained from stereo cameras on top of a vehicle to estimate the ground area of the scene and to define virtual planes. Certain height from the ground area in 2D images is cropped to focus on objects on flat roads. Then, the point cloud is used to filter out false-alarms among the over-detection results generated by an off-the-shelf deep neural network, Mask RCNN. The experimental result showed that the proposed approach outperforms Mask RCNN without pre-processing on a benchmark dataset, KITTI dataset [9].

Format

application/pdf

Number of Pages

42

Publisher

South Dakota State University

Rights

In Copyright - Educational Use Permitted
http://rightsstatements.org/vocab/InC-EDU/1.0/

Share

COinS