Improvements of a retrospective analysis method for a HMM based posture recognition system in a functionalized nursing bed
Julia Demmer, Andreas Kitzig, Edwin Naroska · 2017
In an adverse acoustic environment the human brain can form a sensible sentence only from recognized fragments of a spoken text. It completes parts that are missing and corrects defective parts. This mechanism is called “close from the context”. For a recognition system which basically has no context-related information, the same “skills” would be a very helpful extension. In this work, we show the improvement through an additional recognition module that uses a retrospective process which is able to identify and correct errors in the initial recognition results. The whole retrospective improvement is designed for a Hidden Markov Model (HMM) based recognition system and is realized for a functionalized nursing bed. The recognition system is able to recognize the patient's body posture and actions in the bed and can be used e.g. for sleep pattern analysis or decubitus ulcer prevention. The functionalized nursing bed is designed as a preparation free measuring system which is able to derive biomechanical signals from the patient. For this purpose, it is very important to get a minimum error rate of the used recognition system to avoid false positive or false negative recognition results. The system works as follows. In a first step, the retrospective system analyzes the initial recognition results and classifies them into correct or incorrect results. This classifying is based on probability based distance measure in combination with a context based method. Subsequently the system extracts incorrect marked parts and tries to correct single errors with context based methods initially. In a second step, double errors are corrected by a combination of context based and temporal mean values methods. This work focusses on the development and combination of the two methods to correct double errors to improve the initial system.