Kinect Based Classifiers Design For Hands Images

Pose recognition has been widely used on many applications, as one of thebasic research topics in computer vision area. However, a lot of traditional poserecognition methods are based on some helpful device such as data gloves, or thesemethods use the traditional camera that cannot provide more information than2DImage. Fortunately, the3D-Camera called Kinect brings us convenience incollecting depth images. So that, in this paper, our goal is to build a pose recognitionclassifier platform based on Kinect，on which some basic posture can be easilyrecognized. We use two method on classifying the hands gesture: One is based onRandom-Forest, and in this part, we adopted several traditional scheme feature suchas HOG and LBP; Another method is based on Deep Learning, on which the featurecan learning consecutively all by themselves, and then we brings a classifier such assigmoid on the top of the network to classify the hands gesture data. This papermainly focuses on the following issues:Firstly, we build a pipeline to deal with the video frame collected by Kinectdevice, and then we can get the basis depth images and RGB images;Secondly, the Raw Data we collected by step one is pre-processed according tothe additional information such as the depth of skeleton provided by Kinect, whichmade the experience result more precisely;Thirdly, we designed some feature which are perfectly filled our requirement,and we set appropriate parameters of basic features such as HOG, LBP, with whichthe random forest can be built well;Fourthly, we design another network based on deep learning, on which thefeature can be learning all by itself consecutively, then bring in a classifier on thetop of the network; Finally, we contrast and analyze the different experiment result between thetwo methods.We get a lot of hands gesture data from the consecutive video, and get81.0%recognition result on the test data, and89.17%on the training data, both of whichare based on the deep learning method; however, when we use the random-forestmethod, we got the76.01%accuracy on the test data, and89.01%on the trainingdata.We conclude that both of the two classify method are effective according to thesimilar performance on training data, while the performance based on deep learningis better than that on random-forest method, when we test on the test data, this leadto the conclusion that we can truly get some useful feature based on the deeplearning method.