59. TURBOCHARGING DETECTION CLASSIFIER STATISTICS
Department: Computer Science & Engineering
Faculty Advisor(s):
Ryan Kastner | Kenneth Kreutz-Delgado
Primary Student
Name: Deborah Ellen Goshorn
Email: dgoshorn@ucsd.edu
Phone: 858-353-0914
Grad Year: 2010
Abstract
THE ENHANCEMENT OF LOW-LEVEL OBJECT RECOGNITION CLASSIFIER STATISTICS USING HIGH-LEVEL SEQUENTIAL SYNTACTIC BEHAVIOR CLASSIFICATION
(TURBOCHARGING DETECTION CLASSIFIER STATISTICS)
Multi-class object recognition algorithms exist in all research fields involving any sensors. Such algorithms, termed as low-level classifiers, output object class labels based on sensor data. When such classifiers have mediocre accuracy, the researcher in his or her area endeavors in countless tunings until the classifier performs acceptably for that specific sensor environment. Additionally, it is often the case that the sensor’s environment changes, either by a dynamic environment or the sensor itself being physically moved to another location. In either case, low-level classifiers must be tuned again and often tediously retrained to achieve acceptable performance in the changed sensor environment. A novel approach to eliminate this dilemma is to incorporate a high-level classifier to enhance the performance of the low-level classifier, and thus "turbocharging" the detection statistics of the low-level classifier. This poster presents an innovative high-level classifier which removes such tedious efforts from the researcher and requires no training itself, only a priori statistics of the low-level classifier at the current environment. The proposed high-level classifier is a robust implementation of sequential syntactic pattern recognition which classifies high-level object behaviors, while in parallel recovering from low-level object recognition errors. This grammar-based high-level classifier utilizes weighted costs on the error production rules with a novel method of automating these costs. One of the most difficult sensor research areas for obtaining robust low-level object recognition in dynamic environments is that of computer vision. Thus, to demonstrate the value of using the proposed high-level classifier, this poster shows significant results enhancing the performance of typical computer vision low-level classifiers from very poor to very high (e.g. 33% to 100%) accuracy. The low-level computer vision classifiers attempt difficult multi-class object recognitions such as hand signal recognition, human posture recognition, and, finally, face recognition.
Related Links:
All Content Copyright 2008 Regents of the University of California. All rights reserved.
Official Web page of the University of California, San Diego.
