Unstructured Data Analysis of Streaming Video using Parallel, High-Throughput Algorithms

H. Trease, T. Carlson, R. Moony, R. Farber, and L. Trease (USA)


Unstructured Data Analysis, Video Content Analysis, Face Extraction, Face Recognition, HighPerformance Computing, Video Analysis


This paper describes the use of high-performance, parallel, unstructured data analysis techniques to extract of human faces from streaming video. These faces are placed into a database which may be searched to answer the question “have we seen this person’s face before? If the answer is yes, then where/when have we seen this person and what other faces (i.e., people) were they associating with?” This paper describes the algorithms that are used to: 1) extract all faces from the videos [1], 2) classify and characterize the faces into clusters containing all the different views of the faces, 3) search the face database using a network graphs representing face vs. time, face vs. location and face vs. face to discover social networks/relationships and 4) parallelize the algorithms to achieve optimal use of the processing power and communication bandwidth. We describe the parallel, high-throughput methods (used to process video at a rate of about one DVD (~5Gbytes) per second), and algorithms that extract faces and build attribute vectors computed for each video frame and face based on information entropy relationships. Principal Component Analysis (PCA) [2] is used to identify relationship among the frames and faces to reveal clusters of images and the trajectories of images that define how close one frame is to another or how close one face is to another. Finally, “social” networks are constructed to indicate which faces interacted with which other faces at what location for what period of time. The major use of the system described in this paper would be a scenario where the video feeds of many surveillance cameras are processed into a centralized database.

Important Links:

Go Back