Research And Realization Of Human Activity Recognition Using Kinect

Human activity recognition is an important area in computer vision research,which has great research value and wide application prospects.At present,its applications mainly focus on surveillance systems,patient monitoring systems,human-computer interaction and other fields.The traditional research on human activity recognition is mainly based on RGB image sequences and depth map sequences.The limitations of the research are constantly emerging: both are confronted with complex background interference,and RGB images lose the depth information.The recognition of human activity based on skeleton data,which eliminates the background interference and has the small amount of data,has drawn the researchers' attention.In 2010,Microsoft introduced a somatosensory device called the Kinect that not only captures RGB images,but also obtains depth maps and skeleton information.The advent of this device facilitates the study of human activity recognition based on depth maps and skeleton information.This paper presents a multi-feature fusion method for human activity recognition based on depth maps and skeleton data.For depth maps,using Depth Motion Map-HOG(DMM-HOG)features that are mature and well-behaved.For skeleton data,several improvements have been made to Histogram of Oriented Displacements(HOD)features to create new 3D-HOD features.The post-fusion of DMM-HOG features and 3D-HOD features is used to recognize human activities.The HOD feature uses displacements to construct descriptors,it does not require any preprocessing of the position of the target,and has high computational efficiency.The improved 3D-HOD feature inherits the advantages of HOD features,and improves the defects of HOD features.There are three major improvements to HOD features: 1.Delete the skeleton nodes,remove the hand(Hand Right,Hand Left),the foot(Foot right,Foot Left),spine(Spine),use the remaining 15 nodes for feature extraction,to ensure the accuracy of features while reducing the computational burden;2.The HOD feature uses three projection planes to construct the descriptors of a 3D trajectory.This paper divides the 3D space into 24 directions,describes the trajectory directly in the 3D space,and more accurately reflects the real motion changes,making the descriptor more accurate;3.Utilize the time pyramid that accords with the movement law,make the ideal difference between the low-level histogram in the horizontal direction,more accurately express the time sequence information.In the process of multi-feature fusion,the analysis of the pre-fusion and post-fusion,the advantage of post-fusion is that it can set parameters for different characteristics,makes full play to the advantages of multifeature fusion.For feature selection,using the depth map DMM-HOG features and skeleton data 3D-HOD features fusion.The two features have good complementarity because of the different extraction sources of DMM-HOG features and 3D-HOD features,which makes the description of human activity more accurate after fusion.In the experiments,the MSR-Action3 D data set is used for testing.Multiple configurations of HOD features are compared with the improved 3D-HOD features and which proves that HOD features improvement work is very effective;Multiple recognition algorithms based on depth map or skeleton data are compared with DMMHOG+3D-HOD features,which proves that post-fusion of DMM-HOG features and 3DHOD features has excellent recognition effects.Experiments show that the proposed multi-feature fusion human activity recognition method has good accuracy.