Abstract
Human recognition with skeletal data has the advantage in detecting people without their face characteristic on image. However, the accuracy of recognition by this method is always low because it relies deeply on manual feature selection. We propose a novel human recognition method called Joint Coordinate Images (JCIs) with Convolutional Neural Network (CNN) based on the image generated from skeletal information tracked by KinectV1. In order to represent human physical skeletal characteristic, the coordinate values XYZ of human joint tracked by KinectV1 are firstly created in color image called Joint Coordinate Images (JCIs), in which the relative position of the pixels represents the skeletal structure characteristics of participants with shape in “大” structure. Secondly, a new convolution neural network classifier Lenet-5 model, which always performed well in image classification, was modified to be able to input our JCIs for human recognition. The experimental results show that human recognition using joint coordinate image and Lenet-5 network can reach the highest recognition accuracy of 90.00% on the G3D dataset, which demonstrates the feasibility to transform the skeletal coordinate information into color image for human recognition task and could be used as a complementary method to the well-known application of face recognition.
Human Recognition Using Joint Coordinate Images (JCIs) with Convolutional Neural Network
