基于深度信息和SURF-BoW的中国手语识别算法

摘要

为实现视频中手语的准确识别，提出一种基于深度图连续自适应均值漂移( DI_CamShift)和加速强健特征词包( SURF-BoW)的中国手语识别算法.该算法将Kinect作为手语视频采集设备，在获取彩色视频的同时得到其深度信息.算法首先计算深度图像中手语手势的主轴方向角和质心位置，通过调整搜索窗口对手势准确跟踪；然后使用基于深度积分图像的OTSU算法分割手势并提取其加速强健特征( SURF)，进而构建SURF-BoW作为手语特征并使用SVM识别.通过实验验证该算法在单个手语字母上的最好识别率为99.37%，平均识别率为96.24%.%To realize the accurate recognition of sign language in the video, an algorithm based on depth image CamShift( DI_CamShift) and speeded up robust features-bag of words ( SURF-BoW) is proposed. Kinect is used as the sign language video capture device to obtain both of the color video and depth image information of sign language gestures. Firstly, spindle direction angle and mass center position of the depth images are calculated and the search window is adjusted to track gesture. Next, an OTSU algorithm based on depth integral image is used for gesture segmentation, and the SURF features are extracted. Finally, SURF-BoW is built as the feature of sign language and SVM is utilized for recognition. The best recognition rate of single manual alphabet reaches 99 . 37%, and the average recognition rate is up to 96 . 24%.