HCI (Human-computer interaction) is an important part of people’s daily lives,and the methods of interaction are diverse. Gesture is more in line with the naturalhuman interaction, so that gesture-based interaction has become one of the hotspotsof research in the field of HCI. In order to implement gesture interaction, we needto build a set of gesture recognition system, and the ways of implementation aredifferent. Currently, due to the less restrictions on users, vision-based gesturerecognition has become a main research topic.Gesture recognition system in practical applications, is generally facing the twoproblems: recognition rate and real-time performance. This dissertation mainlystudies the two problems and its main contents are as follows:Firstly, the dissertation describes the main methods of motion detection andproposes a new way of gesture segmentation which is based on motion historyimage and ellipse fitting. Owing to recognition is only done in this movementregion, the scope of searching areas can be narrowed and the amount ofcomputation can be reduced as well.Secondly, this dissertation studies the HOG(Histogram of oriented gradient)feature extraction and the principle of SVM(Support vector machine) classifier, anddescribes the implementation steps of static gesture recognition based on HOG andSVM, and then analyzes the recognition rate and real-time performance of thismethod.Finally, the dissertation puts forward a gesture recognition system combinedwith motion detection. This system can improve real-time performance due to onlyidentifying gesture in the candidate regions. In order to improve the recognitionrate, this dissertation presents two steps: one is comparison of geometric interval,and the other is accumulation of multi-scale detection results. It can be seen fromthe experimental results, the proposed method has reached a high level inrecognition rate and real-time performance. |