UCF Data Systems Group
    |     Search
Projects
  Real-Time Route Diversion System  

The objective of this project is to investigate a computer model and develop a Real-Time Route Diversion System (RTRDS) to study the feasibility of a computer-aided system for efficient route diversion management. The project is currently funded by Florida Department of Transportation (FDOT). The RTRDS is an extensible, adapter-based framework designed to create route diversions, taking into account available real-time and historical traffic information.  Given the need for a route diversion, operators can choose to have the system generate a new route diversion, or select from an archive of historical route diversions. The user interface permits operators to modify generated and historical route diversion plans on-the-fly and then implement them with existing hardware (i.e., dynamic message signs, etc.).  Furthermore, due to the tight integration with Dynasmart-P, operators can save the current route diversion plan and traffic conditions to files that can be used for later performance analysis.

Additional information.

  
   
  SamMatch Image Retrieval  

PROJECT SUMMARY

CONTENT-BASED IMAGE RETRIEVAL

In typical query-by-example systems, users query a database, using images as an example of similarity. Typical images contain objects of interest and also irrelevant image areas, or noise. This noise compromises the effectiveness of retrieval in any similarity model that takes the entire example as relevant; unfortunately, many QBE systems are in this category. For instance, an image is traditionally represented using a few tightly integrated feature vectors, making it difficult to exclude noise in a similarity comparison. Recent approaches aim to specify regions of interest using low-level image features. These attempts are hindered by unreliable segmentation schemes and computationally expensive routines.

In fact there is a solution from the signal-processing domain, in which audio waves are down-sampled and represented in a compressed digitized form. To set up a similar representation for images, a 2-D spatial sampling technique for image retrieval must be developed. In [3], we formalize the concepts involved and provide a unified framework in which the necessary properties for image retrieval (such as translational invariance, scaling, semantic constraints, and no false dismissals) are systematically derived and proved. This work sets out the framework for noise-free query processing in a QBE environment [1][6][7]. The system is supported with various routines, including indexing methods [4] and a query preprocessing function [5].

Concept-based retrieval is the next natural level of content-based image retrieval. A major challenge is accurate representation of the concepts involved. The flexibility of the sampling-based approach allows users to define concepts without hindrance by noise. The proposed framework is again the core component in a methodology for query-by-concept image retrieval [2]. The system automatically annotates images based on a set of noise-free training examples, and enables queries to be formulated using concepts such as tree, sky, and building. Experiments have found that the approach achieves very high accuracy in annotating images.

MULTIDIMENSIONAL ACCESS METHODS FOR MULTIMEDIA DATA

Efficient indexing and retrieval of multimedia data are necessary to a successful multimedia system. Fast response time is critical in the success of many applications such as online multimedia retrieval. Most high-performance indexing techniques are not suited to this task, as the requirements of typical multimedia systems often far exceed the specifications of traditional indexing applications. Today's databases can contain thousands of multimedia objects, rendering impractical a complete scan to find relevant objects. The approximate nature of similarity measures in multimedia retrieval means that large returns are required as a result of inaccurate rankings. A search of ten nearest neighbors (k = 10), as in most research studies, is well below what is needed to offset false dismissals. For large k these techniques are out-performed by a simple exhaustive scan.

The main difficulty, as is well known, is the high dimensionality of multimedia data, which are normally represented by hundreds of dimensions (features). Most indexing methods then break down and perform worse than a sequential scan. This is the notorious curse of dimensionality. It refers to the exponential expansion of the data space that consequently becomes very sparse when populated with a data set of practical size. A direct consequence is that the average distance between neighboring points becomes so large that the standard hypercubic approximation of the search sphere returns almost the entire data set.

We believe that efficient multimedia retrieval should be addressed in the larger context of searching in multidimensional space. We therefore reexamined every aspect of the retrieval issue, including the standard hypercubic approximation. Our detailed analysis in [10] exposes some interesting properties of high dimensional spaces and suggests a promising solution to these problems. In particular, we was led to propose a novel partitioning approach using iso-hypersurfaces, including those generated by the means and standard deviations of datapoint vector components. Advantages of this approach include: (i) the search radius can be larger than the linear size of the space; (ii) the number of dimensions can be greatly reduced, so that existing high-performance indexing methods can be employed; and (iii) the resulting approximation is more accurate than those based on hypercubes. Extensive experiments on multimedia data confirm that this technique greatly reduces dimensionality while improving precision across the entire range of search radii [8]. We are currently expanding this method to speed up the retrieval of multimedia [9].

  
   
  Symantec Project  

We have developed the Backup Grid, a new-generation backup technology for Symantec.  Most existing data backup systems adopt the traditional client-server model. However, this model has the following shortcomings:  static resource binding, no dynamic load balancing, underutilization of some servers, and high administration cost. To address these limitations, we propose a backup-grid approach with the following features: single-instance storage to save storage and communication costs, adaptor architecture to support heterogeneous backup applications transparently, data replication to enhance data availability, data redirection to avoid job resubmission, and intelligent scheduling to balance workload and storage usage among nodes.  Our project has been selected to be presented at the Symantec Cutting Edge Conference 2007.

  
   
  Peer-to-Peer Streaming  

The concept of peer-to-peer (P2P) is not new; however, several factors have lit a fire under the P2P movement today: inexpensive computing power, bandwidth, and storage. Put simply, P2P computing is the sharing of computer resources and services by direct exchange between systems. P2P takes advantage of existing desktop computing power and networking connectivity, allowing economical clients to leverage their collective power to benefit the entire enterprise. This reduces the load on servers and allows them to perform specialized services more effectively. At the same time, P2P can reduce the need for IT organizations to grow parts of its infrastructure in order to support certain services. Realizing the merits of P2P, since two years ago, researchers have applied this concept to many modern applications. Nevertheless, a practical application has received very little attention: media streaming. Given the fact that the Internet currently cannot support IP Multicast while CDN solutions are very costly, P2P could be an efficient and feasible solution to the problem of media streaming. We are interested in finding such a solution.

In a media-streaming P2P architecture, the service delivery tree roots at the server and includes all and only the clients. A subset of clients get the content directly from the server and the others get it from the clients in the upstream. A very critical problem in any media streaming system is the limitation of server bandwidth in serving many clients. P2P alleviates this problem by capitalizing a client's bandwidth to provide services to other clients. However, to design an efficient P2P streaming system requires consideration in the following issues:

  • First, the end-to-end delay from the server to a client may be excessive because the content may have to go through a number of intermediate clients. To shorten this delay, the P2P tree height should be small and the join procedure should finish fast. The end-to-end delay may also be long due to an occurrence of bottleneck at a tree node. Therefore, apart from enforcing the tree to be short, it is desirable to have node degree bounded.
  • Second, the behavior of clients is unpredictable; they are free to join and leave the service at any time, thus abandoning their descendant peers. To prevent service interruption, a robust technique has to provide a quick and graceful recovery should a failure occur.
  • Third, clients may have to store some local data structures and exchange state information among each other in order to maintain the connectivity and improve the efficiency of the P2P network. The control overhead at each client for fulfilling such purposes should be small to avoid exaggerative use of network resources and to overcome the resource limitation at each client.

We recently proposed a solution called Zigzag to address the three issues above. In Zigzag, the multicast peer tree has a height of O(logN) where N is the number of peers, and a degree bounded by a constant. Even though building a tree having such properties is easy, how to efficiently maintain it under network dynamism such as peer leaves and joins is difficult. Zigzag handles this case excellently; especially, failure recovery can be done regionally with impact on at most a constant number of peers and mostly with no burden on the server.

[INFOCOM2003] Duc A. Tran, Kien A. Hua, and Tai T. Do. ZIGZAG: An Efficient Peer-to-Peer Scheme for Media Streaming. IEEE INFOCOM 2003.
 
[SIGMM2002] Duc A. Tran, Kien A. Hua, and Tai T. Do. Scalable Media Streaming in Large P2P Networks. ACM Conference on Multimedia, 2002.
 
[JSAC2003] Duc A. Tran, Kien A. Hua, and Tai T. Do. Scalable Media Streaming in Large P2P Networks.IEEE JSAC Special Issue on Advances in Service Overlay Networks

  
   
© 2008 UCF Data Systems Group All Rights Reserved
Department of Electrical Engineering & Computer Science | University of Central Florida | Orlando, FL 32816
Phone: 407.823.2341   Comments or questions: