In recent years, an increasing amount of urban visual big data is collected through a diverse range of sources, such as taxi vehicle records, video from surveillance cameras, or images captured by mobile devices. The large collection of urban data contains rich implicit information that can help numerous downstream tasks, such as monitoring for construction management companies, planning for government units, etc. However, it is challenging to efficiently extract the desired information from a large-scale dataset. In this work, we focus on developing methods for extracting the spatial attribute and the temporal attribute from urban visual data. Specifically, we introduce a method of organizing large-scale urban visual data into a spatial-temporal structure by mining attributes inherent in the data. We demonstrate the effectiveness of our method by using videos captured by the front-facing camera of buses to detect and analyze work zones within the captured videos. First, the raw set of bus data is preprocessed into a spatial-temporal data structure. Next, we exploit the rich spatial and temporal attributes of bus data in the application of work zone detection and analysis. The goal of this work is to demonstrate the effectiveness of using spatial and temporal attributes to break down large-scale urban visual data and extract insights from large-scale unlabeled data.

46 pages

Thesis Committee:
Srinivasa Narasimhan (Co-Chair)
Christoph Mertz (Co-Chair)
Stephen Smith

Srinivasan Seshan, Head, Computer Science Department
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu