The ThreadMill architecture for stream-oriented human communication analysis applications

Abstract

This work introduces a new component software architecture -ThreadMill -whose main purpose is to facilitate the development of applications in domains where high volumes of streamed data need to be efficiently analyzed. It focuses particularly on applications that target the analysis of human communication e.g. in speech and gesture recognition. Applications in this domain usually employ costly signal processing techniques, but offer in many cases ample opportunities for concurrent execution in many different phases. ThreadMill's abstractions facilitate the development of applications that take advantage of this potential concurrency by hiding the complexity of parallel and distributed programming. As a result, ThreadMill applications can be made to run unchanged on a wide variety of execution environments, ranging from a single-processor machine to a cluster of multi-processor nodes. The architecture is illustrated by an implementation of a tracker for hands and face of American Sign Language signers that uses a parallel and concurrent version of the Joint Likelihood Filter method.

A Distributed and Parallel Component Architecture for Stream-Oriented Applications

“…This tracker is based on a Joint Likelihood Filter multi-target tracker described by Rasmussen [23]. Details of the implementation can be found in [6].…”

“…This application tracks the position of the hands and face of American Sign Language signers from frame to frame of a video input (acquired by Reader). The method consists of evaluating multiple hypotheses based on random samples around positions predicted by kalman filters against evidence provided by a skin-color filtered video frame (details are out of the scope of this paper -see [23] for details of the method and [6] for details of the ThreadMill implementation). Components in Figure 1 are distinguished according to their types.…”