Research On 3D Skeleton Action Recognition Based On Features Fusion

Action recognition plays a key role in applications such as smart security,human-computer interaction,and driverless vehicles.For a long time,RGB images are sources of information for action recognition.With the advent of depth cameras and the development of depth image estimation techniques,a new branch,"3D skeleton action recognition," has emerged in action recognition.The 3D skeleton articulation data is obtained on the basis of depth images by estimation techniques,and has the characteristics of insensitivity to light changes.At the same time,the estimation technique can estimate the spatial position of the blocked human joint points based on the human skeleton structure.In this way,to some extent,the self-occlusion problem that relies on action recognition based on RGB images is alleviated.Due to the above advantages,many scholars have invested in research based on 3D skeleton action recognition.In recent years,after the efforts of researchers,many action recognition methods based on 3D skeleton data have been proposed.Most of these methods use frame-level features to describe poses.However,these features are not robust,and lack of sufficient motion information between frames.Therefore,the action recognition methods based on these frame-level features still have much room for improvement in experimental results.In addition,dynamic time warping is still the first choice for processing sequence matching in many recognition frameworks.However,dynamic time warping is not effective for dealing with asynchronous actions and repeated actions.Based on the above situation,this paper first analyzes the current research status of 3D skeleton action recognition,and then uses the feature fusion method to recognize the skeletal actions in order to obtain a higher recognition rate.The contribution of this article is as follows:(1)This paper presents a MCTD feature.In this paper,special European group SE(3)is used as a feature to describe the rotation and translation of actions,then the coordinate and frame time are fused to locate the spatial and temporal distribution of the action features,and the MCTD(Motion-Coordinate-Time Descriptor)is obtained finally.The MCTD feature represents actions well.(2)The framework of action recognition method based on MCTD features is proposed.First,the MCTD features are calculated,and the similar distances of different action sequences are computed through the MCTD kernel function.Then the High-Order Occurrence Pooling methods are used to mine the high-order autocorrelation statistics of the action features,which solves the problem of the length of action time mismatch and action repetition.(3)A space-pairwise method based on spatial relative distance is proposed to make up for missing spatial structure information.For different actions,the bones involved in the execution are different.In this paper,relative space distances are calculated.By selecting Top-k joint point pairs,trajectories are matched to obtain the space-pairwise descriptor(SPD).Then,the relative spatial distance is computed as a gesture feature to describe the poses,and later fusion operations are performed after that.Finally,the methods described in this paper are applied on three common datasets for experiments.The experimental results show the feasibility and effectiveness of the proposed method.