Crowd Surveillance with the Video Captured by Handheld Devices

摘要

The Gaussian Mixture Model has been widely used to detect foreground in the rapidly changing environment but it fails in low light condition and deteriorates when the background is rapidly changing and it cannot be applied to the video captured by handheld until the frames are stabilized. The proposed work is focused on group detection and people counting in videos where the camera is handheld. We utilize the motion information for each pixel between current and previous frames to prepare a confidence matrix to correctly model background distribution. This work excludes the current distribution of the pixel if the match occurs; in case the pixel is foreground in confidence matrix, reducing the chance of false positive inclusion in background model estimation, and detects around 90% of the foreground pixels. Moreover, the height, width, area of the bounding box of the blobs detected and foreground pixel density are used to distinguish individual and groups, and people counting.