Crowd Density Estimation And Motion Trajectory Detection In Video Surveillance

The research on crowd density estimation and motion trajectory detection is an important part of visual surveillance and crowd management.It is widely used in security monitoring of public places,intelligent management and traffic control.It is an important research topic in computer vision field.Crowd density estimation is mainly based on the pixel and on the texture.In this paper,the threshold segmentation is introduced.Firstly,the foreground is extracted,and the ratio between the sum pixels of foreground and the sum pixels of the background is calculated.Then,when the ratio is lower than the threshold,the crowd is in a low density,and the algorithm is based on pixel.Otherwise,the crowd is in a high density,and the algorithm based on texture is used.For the high density,the paper proposes a method of feature data normalization based on variance,and solves the problem of adaptive normalization.Motion trajectory detection includes target detection,feature extraction,target tracking,and so on.The paper proposes a feature fusion method based on variance regulation for particle filter trajectory detection.Firstly,the HSV feature and GLCM features are extracted.Then,the variance of HSV features and GLCM features are calculated.The weight is large,while the variance is large,and conversely,the weight is small.Finally,the particle filter algorithm is used to achieve the target tracking,and then the trajectory detection is completed.According to update of the feature variance,the feature fusion method based on variance regulation is realized.The method strengthens the characteristics of the bigger influence,and weakens the characteristics of the smaller influence,and it enhances the adaptability of the trajectory detection.In the crowd density estimation,the PETS2009 dataset is selected and the experimental environment is MATLAB R2014 a.The results show that the accuracy of the improved method is 97.5%.In the motion trajectory detection,the PETS2009 dataset and a section of the campus surveillance video are selected,and the experimental environment is Microsoft Visual Studio 2010.The results show that the method has better robustness to color,illumination and occlusion,and it has good effect on the tracking error and stability.