Events Calendar

PhD Defense - Jing Huang

Title: Object Detection and Recognition from 3D Point Clouds

Location: PHE 223

Time: 10 am, Tuesday, Aug 30th, 2016

PhD Candidate: Jing Huang

Committee:

Suya You (Chair), Ulrich Neumann, Aiichiro Nakano, C.-C. Jay Kuo (Outside Member)

Abstract:
Object detection and recognition are fundamental problems in computer vision. While most existing works have been in the 2D image domain, 3D data are gaining popularity in recent years thanks to the development of 3D sensors. My work focuses on object detection and recognition from 3D point clouds, which involves various stages of point cloud processing including feature description, matching, segmentation, localization, classification, and labeling. We explore two different strategies to solve the detection and recognition problems:

The first strategy is to compute a set of descriptors on the local neighborhood of feature points and use them in the matching-based framework. We apply this strategy in the industrial object detection problem, where the intra-class variation is small. We first present a 3D descriptor based on the self-similarity property of the data, and apply the descriptor to build a feature-based matching module. The matching module is incorporated in the object detection and recognition system, which is further used to build an object-level change detection system.

The second strategy is to compute a representation for a whole candidate cluster, and then apply machine learning techniques to classify the clusters without knowing the exact poses. We employ this strategy in the urban object detection, where the intra-class variation is much higher. Specifically, we develop a slicing-based localization method for pole-like objects, introduce a representation of six attributes based on the height and five PCA-based features and apply SVM to classify the candidate objects into four categories. For vehicles, the PCA-based features are not enough to tell them apart from other planar objects. To this end, we employ the deep Convolutional Neural Networks (CNN) based on the orthogonal-view information from the candidates, and prior knowledge for vehicles and urban environment is utilized to help the detection process. Finally, inspired by the success of deep learning on the 2D problems, we present the voxel-based fully-3D Convolutional Neural Network on the point cloud labeling problem. This approach minimizes the use of prior knowledge and hand-crafted features compared to most previous approaches.

We demonstrate the proposed object detection and recognition methods through experiments on point clouds from industrial datasets and large-scale urban datasets.

Location: Charles Lee Powell Hall (PHE) - 223

Audiences: Everyone Is Invited

Contact: Lizsl De Leon