Research On Human Behavior Recognition Based On Spatiotemporal Interest Points

Human behavior recognition is the detection, classification and recognition of human moving objects, it has become a hot research direction in the field of computer vision because of its value and potential in human-computer interaction, video search and other applications. This paper constructs the behavior recognition framework, through the use of local interest points to generate local space-time cube as features, uses multi-dimensional description to enhance the description of the feature, and the improved bag of words model is used to construct the classification model. Finally, the classification results are obtained by training and testing. The main work is as follows:(1) In the stage of feature extraction, the extraction methods and effects of local interest points are studied and experimented.We analyze the popular interest point extraction method of Cuboids extraction method and Harris3D extraction method, select the more efficient and robust extraction method.We establish the interest point extraction method and the parameters of spatial scale and time scale by experiments,so as to determine the size of the space-time cube.(2) In the stage of feature description, this paper proposes two improvements to the LBP-TOP algorithm.? The LBP-TOP algorithm ignores the temporal and spatial differences in the description, this paper proposes a TLBP description method based on time domain characteristics, and it can describe the relations between frames efficiently in the time domain to reflect the variation.? In order to more fully describe the characteristics of information,we put forward a multi dimension description method DT-LBP-TOP for the space-time cube, it can obtain adequate description in time and space, achieve a richer description of video behavior characteristics.(3) On the construction of the bag of words model, the improvement of the original model is put forward.For the original bag of words model, it is easy to cause the problem of single feature partition when constructing dictionary. We propose the use of K-SVD sparse coding to construct a dictionary, which can be divided into several categories to reduce the error of a single partition. Then we focus on the problem of missing the correlation and location relationship between the features, we put forward the lexical tree of Pyramid to describe the structural information of the features, this method enriches the lexical information in the bag of words model.(4) In order to validate the universality of the algorithm, we analyze the behavior of students and design six types' behaviors of students in the process of teaching, we record video data sets validated by the proposed algorithm, and the experiment shows good classification effect, which indirectly proves the application potential of this algorithm.