Research On Key Technologies Of Hand Gesture Recognition Based On Depth Information Under Complex Scenarios

With the growth of Artificial Intelligence,hand gesture recognition,as an important technology in human-computer interaction,gradually plays a critical role in human life.Currently,depth information is commonly introduced in hand gesture recognition with the industrial development of 3D vision sensors.However,many problems still exist in hand gesture recognition based on depth information.These problems are concentrated in hand segmentation and feature extraction,including low robustness of hand segmentation,poor quality of feature extraction,heavy computation burden,etc.In this paper,we proposed a novel hand segmentation and feature extraction algorithm based on depth information.The hand segmentation process is on the basis of threshold method which combines background elimination and depth histogram,along with image preprocess and plane detection to enhance the adaptation of various scenarios.The feature extraction process is centered on contour optimization based on Mean Shift method to increase the quality of hand contour.Than it combines the Hu invariant moments of both hand contour and convex hull as a new feature vector.The algorithm is evaluated in terms of computing speed,scenario adaptation and segmentation accuracy under four predefined scenarios including micro spur scenario,family scenario,vehicular scenario and public environment scenario.Eventually,the proposed algorithm is separately cascaded to SVM classifier and CNN classifier to generate two hand gesture recognition systems which provide high accuracy and less complexity.Compared with other algorithms,the proposed algorithm can provide real-time(at most 33.1fps)hand segmentation and feature extraction with better scenario adaptation and higher segmentation accuracy for Kinect2.0 images(512 x 424 resolution)and EPC660 images(320 x 240 resolution)on normal PC platforms(Intel Core i7,without GPU).The classification accuracy is further increased in hand gesture recognition systems based on SVM and CNN to 90% and 92%,respectively.