The Enhancement of Low-level Classifications in Sequential Syntactic High-level Classifiers

This paper surveys a new research field of object behavior classification using sequential syntactic pattern recognition, which recognizes high-level object behaviors while in parallel recovering from low-level object recognition classification errors. A new approach of syntactical object behavior classification with a robust implementation is introduced. It is an innovative approach that requires no training, only a priori statistics of the low-level classifier. This new approach also utilizes weighted costs on the error production rules with a novel method of automating these costs. Other high-level object behavior classifier models, such as variants of hidden Markov models, which attempt to detect high-level object behaviors while recovering from low-level object classification errors are also reviewed. Approaches are explained using one application in mind: hand gesture behavior recognition of temporal sequences composed of hand posture labels; these hand posture labels are outputted from a low-level hand posture recognition classifier with high hand posture misclassification rates. Finally, results are shown from applying the introduced novel approach to hand gesture behavior recognition.