| Summary: | 碩士 === 大同大學 === 資訊工程學系(所) === 99 === This thesis presents a human action recognition method based on Hidden Markov Model (HMM). Two features of the shape contour based histogram and skeleton are extraction and merged into a new feature vector to description a human action. In our propose method, one set of time-sequential images is converted into a sequence of image, and the sequence is transform into a symbol sequence by Euclidean distance. We design a codebook which contains each defined action type and compute the similarity between feature vectors. Each feature vector of the sequence is matched against the codebook and is assigned to the symbol which is most similar. By this way, time-sequential images are transformed into a symbol sequence. We use HMM to model each action type. In the learning phase, the parameters of HMM are optimized so as to best describe to the training sequences. For action recognition, the model which is best match with the sequence is chosen as the recognized type. The experimental results show the effectiveness for action recognizing action. A 92.9% recognition rate is obtained.
|
|---|
