Research On Feature Detection And Description Algorithm Based On Video Appearance And Motion Information For Behavior Recognition

In this paper,we focus on an important topic in computer vision,namely feature detection and description in video.In order to fully explore and utilize the appearance and motion change information in video,this paper uses geometric algebra,a simple and efficient mathematical tool,to do the in-depth research on the feature detection and description algrorithms of video images The main research contents are as follows:1)A new speeded-up robust feature(SURF)detection algorithm based on a unified model of video appearance and motion-variation information(UMAMV-SURF)is proposed.Firstly,using the geometric algebra as the mathematical framework in this paper,based on the model of video appearance and motion information in geometric algebra.the integral video on the UMAMV model is calculated,then the Hessian matrix in the integrated video under the UMAMV model is constructed to obtain the local maximum.In order to simplify the convolution operation of constructing Gaussian second-order differential and integral video in Hessian matrix,In this paper,a three-dimensional box filter is designed for the UMAMV model.The box filter and integral video can simplify the convolution operation of Gaussian second-order differential and video into simple addition and subtraction,which greatly reduces the computational complexity of the algorithm,and builds the scale space on the UMAMV model by changing the size of the box filter.So that the extracted feature points have good scale invariance.Finally,The feature points of this paper are extracted by the method of 3D non-maximum suppression.And the experimental results also show that the feature points extracted by the UMAMV-SURF algorithm proposed in this paper can not only reflect the local information on the video space domain but also reflect the motion information in the time domain.2)A new SURF description algorithm based on a unified model of appearance and motion-variation information(UMAMV-SURF)is proposed.Firstly,the Haar wavelet response value is calculated in the space domain and time domain directions of the feature point neighborhood on the UMAMV model,then the main direction of the feature point on the UMAMV model is calculated,And the feature point coordinates and the response value are rotated according to the main direction of the feature point to ensure the direction invariance of the feature point.Finally,the description vector of the feature points is obtained by statistically calculating the response values of Haar wavelets on the UMAMV model and normalization.The experimental results show that the UMAMV-SURF feature description algorithm proposed in this paper not only reduces the computation time of the algorithm,but also improve the video behavior recognition accuracy.3)A new Two Stream Inflated 3D Conv Net based on a Unified Model of Appearance and Motion-Variation information(UMAMV-TSI3D)is proposed.Firstly,the motion information in the UMAMV model is decomposed into motion components in both the x and y directions and composed into two independent picture channels.Then it combines the apparent information and the information of decomposed motion in UMAMV model with the I3D(Inflated 3D Conv Net)network to form the apparent information I3 D network and the motion information I3 D network of the in UMAMV-TSI3 D network.The two networks train their respective models and obtain their respective predictions and the results of the fusion of the predictions.The experimental results show that the network proposed in this paper can effectively improve the video behavior recognition accuracy.