Optical Flow based Crowd Counting in Video Frames
Sayanti Bandyopadhyay · 2019
The aim of this paper is to count crowd in video frames. We propose to detect the crowd by focusing on the stationary and moving people in the crowd separately. A deep learning based model is trained on the crowd videos for the stationary and moving part. Features are extracted in two different pipelines (i) from each pair of consecutive frames in a video and (ii) from the optical flow between each pair of frames using CNN. The feature maps obtained from the two pipelines are combined at feature level by mapping them to a single density map. The density map has been used for detecting crowd in video frames. Use of the features from both the pipelines have shown to perform well in crowd detection and counting.