New version page

Forensic Video Reconstruction

Upgrade to remove ads

This preview shows page 1-2-3 out of 9 pages.

Save
View Full Document
Premium Document
Do you want full access? Go Premium and unlock all 9 pages.
Access to all documents
Download any document
Ad free experience
Premium Document
Do you want full access? Go Premium and unlock all 9 pages.
Access to all documents
Download any document
Ad free experience
Premium Document
Do you want full access? Go Premium and unlock all 9 pages.
Access to all documents
Download any document
Ad free experience

Upgrade to remove ads
Unformatted text preview:

Forensic Video ReconstructionLarry Huston†[email protected] Sukthankar†‡[email protected] Campbell†‡[email protected] Pillai†[email protected]†Intel Research Pittsburgh‡School of Computer Science417 S. Craig Street Suite 300 Carnegie Mellon UniversityPittsburgh, PA 15213 Pittsburgh, PA 15213U.S.A. U.S.A.ABSTRACTThis paper describes an application that enables quick reconstruc-tion of interconnected events, sparsely captured by one or moresurveillance cameras. Unlike related efforts, our approach does notrequire indexing, advance knowledge of potential search criteria,nor a solution to the generalized object-recognition problem. In-stead, we strategically pair the intelligence and skill of a human in-vestigator with the speed and flexibility of a parallel image searchengine that exploits local storage and processing capabilities dis-tributed across large collections of video recording devices. Theresult is a system for fast, interactive, brute-force video searchingwhich is both effective and highly scalable.Categories and Subject DescriptorsH.3.3 [InformationSystems]: Information Storage and Retrieval—Information Search and Retrieval; C.3 [Computer Systems Orga-nization]: Special-Purpose and Application-Based Systems; I.4.9[Computing Methodologies]: Image Processing and ComputerVision—ApplicationsGeneral TermsAlgorithms, Human Factors, PerformanceKeywordsvideo retrieval, active storage, interactive search1. INTRODUCTIONAs surveillance cameras proliferate, the resulting glut of videoposes a crippling data interpretation challenge. Present approachesto surveillance video analysis rely on linear searches performed byPermission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.VSSN’04, October 15, 2004, New York, New York, USA.Copyright 2004 ACM 1-58113-934-9/04/0010 ...$5.00.human beings to interpret and correlate observed activity betweenmultiple sources. This technique rapidly becomes intractable as thenumber of deployed cameras increases. For forensic reconstructionof events, such as crime scenes, these approaches will always betoo slow to provide real-time guidance to investigators in a rapidly-developing situation.Researchers in a number of fields are pursuing a variety of par-tial solutions to the above problems; many commercial surveillancevideo recorders utilize extreme temporal and data compression toreduce storage requirements and search time, but in doing so dis-card potentially important video frames from the outset and pro-vide no alternative to linear human search. Sensor network ap-proaches [9,12,13] address the question of adapting a collection ofdata gathering nodes or cameras to look for particular patterns, butare continuous query systems which do not archive raw source data(video) nor provide facilities for ex post searching except as antic-ipated by prior queries. Content-based image retrieval [2, 5] andobject detection techniques [18, 20] offer a number of potentiallyinteresting domain-specific classifiers and recognizers, though noalgorithm offers a general resolution to the object recognition prob-lem, and most algorithms are too computationally expensive to runon every frame of surveillance video (e.g., the Blobworld [2] im-age segmentation algorithm takes more than a minute to processa single frame on modern hardware). Image indexing in general,whether indices are built by machines or by humans, can offer fastresponse to queries, but only by increasing the cost of video ac-quisition and only when potential query criteria can be identifiedin advance. Recent research in video surveillance and monitoring(VSAM) [3,7] addresses online image understanding, object recog-nition, and tracking. In this paper we focus instead on searchingraw video data. Any advances that add semantic information to thevideo data as it is captured are complementary to our approach, andthis additional information has the potential to make our searchesmore efficient and fruitful.We propose a new approach to surveillance video analysis whichcombines the interactive use of a human investigator’s skill withthe speed and flexibility offered by the automated, highly parallel,brute-force application of image processing techniques. Interac-tion with a human investigator is essential because fully-automaticextraction of semantic content from images remains significantlyinferior to human performance, despite decades of research in com-puter vision and image processing. For the foreseeable future, theuser should be a key component of any forensic video reconstruc-tion system. Automatic image processing could best be used toeliminate clearly irrelevant data, reducing the amount of informa-tion presented to the user, and utilizing the human investigator’slimited attention more effectively. Large-scale parallelism is re-quired in order to permit quick searching of a large and grow-ing number of cameras and to enable a distributed search to exe-cute close to data storage locations where it will be unhindered byWAN/MAN/Wireless bandwidth limitations. Finally, brute-forcesearch (as opposed to indexing) is unavoidable in surveillance videoanalysis because search criteria will generally not be known untilafter an incident has been recorded. We have implemented our ap-proach in VideoFerret, an application that leverages recent work inautomated, highly parallel, brute-force image search [10]. This isintegrated into a user interface that emphasizes flexibility, concur-rency, fast response times, early presentation of partial results, andspatial and temporal visualization techniques.This paper’s novelty lies not in its image processing algorithms,nor in any efficient scheme for indexing the data, but rather in theability to perform user-defined queries on unstructured surveillancedata and to visualize events through time and across multiple cam-eras. Our approach to the forensic video reconstruction problemenables both human and computer resources to tackle aspects thatare best suited to their abilities: the user can focus on interpretingand correlating semantic information while the


Download Forensic Video Reconstruction
Our administrator received your request to download this document. We will send you the file to your email shortly.
Loading Unlocking...
Login

Join to view Forensic Video Reconstruction and access 3M+ class-specific study document.

or
We will never post anything without your permission.
Don't have an account?
Sign Up

Join to view Forensic Video Reconstruction 2 2 and access 3M+ class-specific study document.

or

By creating an account you agree to our Privacy Policy and Terms Of Use

Already a member?