[C3] Conference Paper
Authors:
A. Roy, N. Biswas, S. K. Saha and B. Chanda
Venue:
Proc. of 2019 IEEE Region 10 Symposium (TENSYMP)
Date:
7-9 June, 2019
Abstract:
In this work a simple methodology has been proposed to categorize video sequences of moving crowd. Based on the motion pattern such crowds are classified either as structured or semi-structured or unstructured ones. Interest points detected in the first frame of the sequence are tracked over the sequence using optical flow. Thus, it requires tracking of only a subset of points in the frame. Based on the motion orientation of such tracked points descriptor is computed. By concatenating the block level histograms of motion orientation frame level feature has been computed. Thus it can well capture the localized motion patterns present in the segments of crowd. Frame level features are concatenated to represent the sequence. Finally, a neural network with multiple hidden layers has been used to classify the sequences.
Links: