Study On Face Recognition Based On Kinect

Human face images acquired using conventional 2D cameras have inherent restrictions that hinder the inference of some specific information in the face.In order to achieve better performance,3D faces captured via specialized acquisition methods have been used to develop improved algorithms.While such 3D images remain difficult to obtain due to several issues such as cost and accessibility,the low-cost depth sensors such as Microsoft Kinect introduced in late 2010 allow extracting directly 3D information together with RGB color images.This provides new opportunities for computer vision and face analysis research.The main works can be summarized as follows.1.A series of face recognition methods based on RGB-D data are investigated.We explore the usefulness of the depth information acquired by the low-cost depth sensor,Microsoft Kinect in the face recognition task.And we confirm a series of conclusions.Depth information alone provides promising classification results beyond the expectations based on human perception.Combining the RGB images with the depth information does provide performance enhancement and the performance of Kinect depth images highly depends on the processing step which is a very crucial step before further analysis.The above proves the significance of studying RGB-D face recognition.2.A compact binary feature for RGB-D face description and recognition is proposed.Firstly,different from traditional hand-craft feature,we learn the compact binary feature from the training set using unsupervised learning method,which can automatically make a compromise between invariant and distinction.Then,in order to make full use of the contextual information,we use the pixel difference vectors as the input instead of using the original training set.Finally,considering the smoothness of the depth image compared with the RGB image,we extract different size of pixel difference vectors from every block of RGB and depth image.This work demonstrates that the proposed method is highly discriminable with good performance in face feature expression and is robust to facial occlusion and illumination.And recognition rates are comparatively high on two publicly available RGB-D Kinect database.3.A depth enhancement method for Kinect v2 and a face detection method based on depth image for real time face recognition with Kinect are proposed.Firstly,we train random forests with the depth images labeled with head locations and orientations,because the relationship between the size of the face and the distance of the face belongs to a second order polynomial,so we can crop the face.Then,we train seven kinds of active appearance model.The estimated location and orientation of a person's head is used for the initialization of the Multiview active appearance model and model choosing of Multiview AAM under depth data,which helps to quick and accurately localize the facial features.Finally,we track the face and extract the compact binary feature for face recognition.Compare with traditional face detection method,the detection rates are comparatively high on the conditional that the environmental light is weak or the head has large pose changes.The real time face recognition method obtains a good performance.