Research On Human Action Recognition

Human action recognition has been one of the most challenging research fields of the computer vision,which has been the current hot researches and already drawn the interest of many researchers.Besides,there are wide and promising application areas in human action recognition,such as,smart video surveillance and security,human-computer interactions and motion action analysis.However,with the advent of the aging society,the falling-action of old people occurs from time to time.The old is likely to be dead if he or she can't be rescued timely after fell.Therefore,the detection of falling-action has great social significance.The central point of this paper for human action recognition is the identification and classification of falling and crouching,bending,sitting and so on.The research on recognition of falling-action and other behaviors in this paper is conducted in the following aspects: moving object detection,detection and removal of moving target shadow,design and implementation of falling-action detection algorithm.The hybrid Gaussian modeling and vibe algorithm are implemented to detect and extract the moving object,finally the vibe is used in motion target detection due to its higher effectiveness and better performance of extracting integrated moving target.There are shadow areas which have the similar property of motion with moving object in the circumstance of lighting and it has a great impact on feature extraction and action recognition.Therefore,we compared the methods of RGB color space,HSV color space and the combination of gray image processing and LBP feature.It is showed by the experiment that the shadow areas can be removed by the combination of adapted threshold gray image processing method and LBP feature.In the aspect of design and implementation of falling-action detection algorithm,the ratio of the height and the width of the moving object is used to distinguish whether the current behavior may be the falling-action or not.Then,the Hu moments are extracted and employed to train and classify by SVM,moreover,we also use convolutional neural network to detect falling action and non-falling action,the image is regarded as the input of the network structure,which extracts the feature by the computer.The algorithm extract the local feature by convolution and pooling operation and then these features are aggregated into global feature,therefore,the image can be represented by these features.It is proved by the experiments that the algorithm studied in the paper has good performance and practicability for the detection of falling-action.