Related Issues Of Feature Extraction And Representation In Behavior Recognition

Tom M.Mitchell said it is one of the most important factors that whether or not machines are given the ability to recognize active behavior and it is because that the given ability makes the machines interact with people and things directly and naturally without specific hardware interfaces.Nowadays,mankind has been standing in the tide of artificial intelligence trends.It is no doubt about application of artificial intelligence,but it is a great challenge how to efficiently use the technology of artificial inteligence.Improving the capability of autonomic behavior recognition is critical solution for the above problems.Behavior recognition is one of the hot and difficult researches in the multi-disciplinary field.It is critical for recognizing and understanding behavior correctly through extraction and representation of behavior feature efficiently,and a lot of research results have been achieved.Due to the differences of research objects and facing problems in various scene applications,there are many problems are as follows in feature extraction and description.Firstly,the existing methods are less robust,easy to be affected by noise,inaccurate in feature carrier detection,etc.In addition,there are the loss of feature diversity including global context information and color texture information of the object description.Therefore,this paper focuses on the above issues and the main contributions are as follows.(1)We proposed some novel qualitative trajectory calculus for abstracting and representing of trajectory feature which were be used for Complex interactive behavior recognition.QTC is easily affected by the negative factors such as noises,the loss of global context information and so on.And then in this paper,we proposed three novel qualitative trajectory calculus,point-point QTC,sub-sub trajectory QTC and point-trajectory QTC,which were be used to contain the more detail information of trajectory,smooth noises and reserve the more global context information.Through testing on the simulate interactive trajectory datasets proposed by us and the traffic datasets,the results show that our novel calculus are better than QTC and the other classical works based on the trajectory-based feature.(2)In this section,we proposed a novel discrimination model of local activity to detect and locate the active soccer players in the sport videos which is use for feature abstraction and representation of tactical event.Most current works of trajectory-based feature through multi-target tracking are affected by the negative factors such as target occlusion and poor identification.To solve the above problems,we proposed a novel discrimination model of local active degree to detect and locate the active soccer players in the sport videos and built a non-trajectory framework of tactical event recognition.The basis realization of the framework is as follows.Firstly,court line detection and camera calibration are carried out to rebuild the three-dimensional coordinate system and segment the field.Secondly,the proposed discrimination model of local active degree is used to detect and locate the active players in the video and then model the spatial-temporal distribution of active players with the segmentation of the field for tactical event representation and recognition.Finally,the proposed framework is used to recognize the attack tactical events of the 2012 European Cup and the Barcelona games in 2013/14 season of the Spanish Football League.The experiment results show that the proposed method is more efficient than the citation methods.(3)In this section,An adaptive Local binary patterns on three orthogonal planes is proposed to abstract the feature and represent in surveillance videos.Currently,the surveillance video of farming environment is susceptible to noises and there are few studies on complex behavior recognition of livestock.Therefore,based on the idea of nonlinear smoothing noise using median filter we proposed an adaptive Local binary patterns on three orthogonal planes operator on the basis of the classical LBP-TOP operator.The new operator adaptively adjusts the central threshold with the linear weighted values of the neighborhood pixels of the window,so as to realize smoothing video noises and reduces the influence of noise on feature coding.Meanwhile,we proposed a simple framework to recognize cows'.Firstly,we used spatio-temporal interest points and ALBP-TOP operator to extract and represent the character of cows' behavior.And then,the bag of feature is constructed using extracted features for basic behavioral recognition of cows.Finally,through the observation of the occurrence frequency of the basic behavior,we revealed the law of prenatal behavior.(4)A novel method of object segmentation is proposed by us which is name the pulse coupled neural network based on a modified artificial bee algorithm.In video frames,it is very important for behavior recognition and understanding to extract the contours and regions of moving boundaries through object segmentation accurately.Most of the current methods of object segmentation act on binary images or grayscale images,which lost amounts of color texture information.Therefore,an improved artificial bee colony algorithm is proposed to optimize the parameters of pulse coupled neural network for object segmentation.Firstly,we proposed an improved artificial bee colony algorithm by adjusting adaptively the search strategy through introducing scale factor and using the entropy weighted linear function as the profit and loss evaluation function.In additional,the optimal PCNN based on the proposed MABC is applied to object segmentation of RGB color image.The experimental results show that the segmentation appearance is more subtle and the color texture information is better remained in the target area based on the proposed method.In summary,this paper focuses on the related issues of feature extraction and representation in behavior recognition,and proposes solutions and experimental verification.The related work of this paper provides a reference for future research.