Human Action Recognition Based On Depth Maps

Human action and activity recognition in computer vision and pattern recognition and other cross areas has always been a popular research topic,which has a broad application prospects such as human-computer interaction,intelligent home,intelligent monitoring.The early human action recognition research mainly focuses on the video sequences taken by visible light camera,and researchers put forward many classic algorithms,but because of the inherent limitation of this type of data,these algorithms are insensitive to change in lighting conditions.With the development of the sensor technology,the appearance of cost-effective depth cameras,such as Kinect,provides new possibilities to address difficult issues in human action recognition.The depth data not only don't easily influenced by factors such as illumination change,it can also provide additional the 3D structural information of the scene.Thus more and more researchers are motivated to explore approaches based on depth data.In this paper,we focus on recognizing human actions from the original depth data and realize a variety of efficient human action recognition methods.The concrete content are as follows:Firstly,based on the depth maps,we proposed an action feature description method.The time sequence is divided by using the adaptive depth motion map energy.The motion energy model(MEM)is obtained by analyzing the motion cues of different sub-periods.Local binary patterns(LBP)is used to gain the MEM-LBP descriptor to represent the action characteristics.After reducing dimension by principal component analysis(PCA),it is sent into the cooperative representation classifier for action recognition.Finally,the experimental analysis and comparison on MSR Action3 D and MSR Gesture3 D datasets demonstrate effectiveness of the proposed approach.Secondly,the multiple fusion method is introduced.Based on the MEM-LBP descriptor,the histogram of oriented gradient(HOG)is extracted to obtain a new feature descriptor,and two types of fusion consisting of feature-level fusion and decision-level are used to combine two feature descriptors,the extreme learning machine(KELM)is used to classify the actions.The experimental evaluation on the datasets show that the two types of fusion methods improve the action recognition results to some extent,moreover the approach of decision-level fusion outperforms the other one.