People Counting Based on Feature-Regression and Detection

Aiming at the deficiency of two main statistical methods of feature regression and detection, the paper presents an approach to combine the two methods legitimately. For prospective area of the video frame, the background segmentation method is used to separate the foreground blocks and features are extracted to estimate the number of people by Bayesian multiple kernel support vector regression; For close shot area, HOG features are extracted to train the models by weak-label structural SVM, then cascade detection with star-structural models to achieve the accurate location and statistics of pedestrians. Experimental shows that the proposed method is not only able to achieve the traffic statistics of people more accurately and reduce the statistical time to some extent, but also can determine the locations of pedestrians in a certain scale.