基于关键帧的多级分类手语识别研究*

提出了一种基于关键帧识别的多级分类的手语识别方法，该方法采用HDR(多层判别回归)/DTW（动态时间规正）模板匹配多级分类方法。根据手语表达由多帧构成的特点，采用SIFT(尺度不变特征变换)算法定位获取手语词汇的关键帧，并提取其特征向量；根据手语词汇的关键帧采用HDR方法缩小搜索范围，然后采用DTW比较待识别的手语词特征与该范围内每一个手语词进行匹配比较,计算概率最大的为识别结果。这种方法在相同识别率的情况下比HMM识别方法速度提高近8.2%,解决了模板匹配法在大词汇量面前识别率快速下降的问题。

Key frame based multi-level classification of sign language recognition

This paper presented a sign language recognition method based on the multi-level classification of key frame recognition. This method adopted hierarchical discriminant regression (HDR) and dynamic time warping (DTW) template to match multi-level classification. According to the multi-frame characteristic of sign language, adopted the scale-invariant feature transform (SIFT) algorithm to orient and obtain the key frames of sign language vocabularies, and extracted the feature vectors. Based on these key frames of sign language vocabularies, the adopted HDR method could narrow the search scope. Then used the DTW compare the irrecognition features of sign language vocabularies with every sign language word inside this scope, and the maximal calculate probability was the recognition result. With the same recognition rate, this method could be 8.2% faster than the HMM recognition method, and solved the problem that the template matching was suddenly slow down in the face of a large vocabulary.