High Performance Gesture Recognition Using Probabilistic Neural Networks and Hidden Markov Models

Publisher Summary This chapter presents a fast method to recognize video sequences with statistical methods without implementing special rules depending on the content of the sequence. The system could learn to classify a test with good results for a dependent person as well as for an independent person while ignoring great parts of the information included in the movie sequences, although classification of the given gestures doesn't seem to be trivial, especially for the gestures NOD-YES and NOD-NO where only the head of the person is moving. The results show that the feature extraction works quite well. The chapter observes the tendency of superior performance for the Markov models compared to the neural net approach. The number of samples is increased to make the results statistically more reliable and the recognition more robust. With improved feature extraction in the preprocessor the system proposed could be a good base for complex future applications, classifying maybe hundreds or thousands of sequences for person-independent tasks in real time.