Human Action Recognition Based On Target Detection

As an important part of human motion analysis,motion recognition has been widely used in daily life,such as human-computer interaction,video surveillance and so on.Because of the non-rigid characteristics of human actions,the meanings of various actions in different environments or conditions are totally different.Because of the difference of the meaning of the same action expression or the same or similar meaning of the different action expression,the difficulty of human motion detection and recognition is often much greater than that of other target detection and recognition.This thesis first detects the human target in the video,and then analyses and processes the target to achieve the effect of classification.The main problems in this thesis include:(1)target detection and action recognition classification in a single environment;(2)target detection and action recognition classification in a relatively complex environment;(3)action recognition classification in static pictures where objects interact with human bodies.Starting from the above problems,this thesis has made corresponding improvements,mainly including the following aspects:(1)A human motion recognition method based on improved spatio-temporal interest points is proposed.This method uses Harris-Laplace spatio-temporal interest points to detect objects in order to obtain relatively drastic changes of pixels in human motion video.Then it is described by 3D-SIFT descriptor and clustered by K-means to form word bag.Finally,multi-class support vector machine(SVM)is used to recognize and classify human motion.In this part,through the experimental verification of KTH dataset,we can see that the redundant points in the detected interest points have been eliminated,and the background points have been suppressed.(2)A human motion recognition method based on improved SSD algorithm is proposed.This method uses SSD detection method to detect targets and background information of images,so that the network can take into account the dual information of targets and background in a forward transmission,thus improving the classification accuracy.In the aspect of action recognition and classification,the network structure is improved,and the traditional convolution in the network model is changed into separable convolution,which not only reduces the experimental process.The number of parameters also reduces the computational complexity of the experiment.(3)A human motion recognition method based on GoogLeNet is proposed.This method improves the GoogLeNet model by using the idea of transfer learning,which enables the network to have certain postural expression ability for the types of individual behavior after pre-training;uses the logical regression multi-classification in logical classification to achieve multi-classification of actions,and verifies it by establishing an action recognition model application system.In this part,the model is tested by MATLAB2017 processing platform,and the average recognition rate of the test image set is obtained.