Video Behavior Recognition Based On Incremental Deep Learning

With the popularity of personal smart equipments and the large-scale application of ‘Sky Net’ surveillance system,acquisition of video data becomes more and more convenient,and thus there are more and more potential applications of video data analysis and processing,including automatic video surveillance,human-machine interfaces,sports video analysis and video retrieval.Human action recognition is the core of behavior analysis and the basis of intelligent video analysis.Its task is to extract the features from the obtained video data,and then classify the actions inside these video samples.In recent years,most contries in the world attaches great importance to public safety,and how to classify different actions in the surveillance video data is thus becoming a hot spot in the field of computer vision.Acquisition and processing of video data are susceptible to many factors,such as the changes of the natural environment in the background of video data,the ambiguity of different actions,the changes of the body’s appearance,etc.,which will affect the reliability and precision of video action recognition.By analyzing the existing framework for video action recognition,it can be found that if the application scenario can be limited to a low complexxity background environment such as indoor monitoring,the problem can be simplified using motion history images.In this thesis,we mainly study how to use deep networks to extract features and combine them with high efficiency classifiers to classify the actions in video samples.Firstly,the advantages and disadvantages of the two main aspects of video action recognition in traditional structures are analyzed,which are feature extraction and classification.The traditional method of video feature extraction based on motion history image and histogram of gradient is tested by experiments,and the advantages of kernel nonlinear classifier in recognition rate and efficiency are also verified by experimental results.Secondly,we study two typical deep learning algorithms,stack sparse auto-encoder and convolution neural network,and apply them to the feature extraction from motion history images of video samples respectively.Further,we use the transfer learning strategy to deal with the problrm of long training time in the deep networks and extract features in small sample conditions using the deeper networks.Experimentsal results show that the deep learning algorithms perform much better than the traditional feature extraction method.At the same time,a kernel based nonlinear representation is combined with the deep networks to improve the efficiency of classification under the premise of ensuring the recognition rate.Finally,by studying a non-negative sparse auto-encoder algorithm,an incremental learning method is combined with the deep learning algorithms so that the features will be learned continuously,and thus the long training time problem is solved from another aspect.