Study On The Method Of Fine-grained Action Recognition In Video

Recently,fine-grained action recognition has gained a lot of interests due to its potential application in human-computer interaction,monitoring security,intelligent robot,and other scenes.Fine-grained action mainly occurs indoor environment,the changes of the motions are mostly centered at terminate joints,with small movement range and complex interaction between human and objects.The position of hand,head and other important parts of human body can be determined by human pose data,which can reflect the subtle changes of joints.Therefore,this paper is based on the human pose to recognize fine-grained actions.The main work is as follows.A pose data smoothing algorithm based on joint confidence is proposed to solve the noise problem of pose data.Firstly,the poses of the same person in the video are divided into sequences.And then the joint points in the pose sequence are smoothed with the joint confidence weight to suppress the noise in the trajectory of the joint points.Joint features and limb features based on the graph structure are proposed to enhance the expression ability of the model for fine-grained actions.In this paper,the object is generalized into node and is combined with human body pose,virtual connection is established between the object nodes and hand nodes,and the structural features of the joint graph of human object interaction are constructed.Limbs can transmit more obvious spatial change information,and the limb orientation information is extracted from pose data to construct limb graph structure features.A two-stream graph convolutional network is proposed for fine-grained action recognition.The two-stream network modeling joint features and limb features respectively,and express actions from different aspects.In the two-stream graph convolutional network,the feature channel weighting mechanism is introduced to highlight the fine-grained information in the pose data.Meanwhile,the classification scores of two network branches are fused in the output layer to get more accurate recognition results.Experiments in Kinetics dataset and MINTA dataset show that the noises in pose data are reduced after smoothing.The human-object interaction features based on the joint graph structure can improve recognition degree of fine-grained action.The feature channels weighted algorithm can enhance the model performance and the effect of recognition in two-stream network get a better result.