An Automated Image-Based Framework For Tracking Pedestrian Movements From Top-View Video
Loading...
Date
2017-08
Authors
Md Yatim, Halimatul Saadiah
Journal Title
Journal ISSN
Volume Title
Publisher
Universiti Sains Malaysia
Abstract
Nowadays, pedestrian safety and monitoring is a very important
consideration in many situations. Detecting, tracking and extracting pedestrian
movements from video can be used to gain more understanding on crowd features
and behaviours. Beside using the extracted pedestrian data for security purpose,
these data could also be useful as an empirical data to calibrate with a simulation
model, enhancing design architecture and also alert the security personnel on
anomaly of events. The existing works on pedestrian detection and tracking have
some limitations. For example, some works focus on a specific event and a
specific place, and some require a lot of human intervention. Therefore, an imagebased
framework for detecting and tracking pedestrian movement from a topview
video is proposed. By using the top-view video, it is possible to allow the
detection and tracking of pedestrian to be done automatically or with no human
intervention. The proposed framework consists of several steps namely the
pedestrian detection and tracking, and image calibration. In pedestrian detection,
the two steps which are feature categorisation and feature grouping are proposed
to be added in the standard detection method. In pedestrian tracking, the hybrid
technique of point tracking and feature tracking is proposed. In order to calibrate
images extracted from video, an algorithm to normalize the pixels is presented.
This framework extracts the trajectory, estimated speed of the pedestrian and
number of pedestrians. This automated framework can be used in some specific
places such as at entrance of a building or a hall. The validation, evaluation and
analysis of the framework is conducted on two case studies namely a set-up
experiment in a controlled environment and an experiment in a non-controlled
environment. Detailed results of the framework and the analysis show that the
proposed framework detect the exact number of pedestrians, correctly plot
trajectory of pedestrian and the percentage of error for the difference of
calibrated, uncalibrated speed measurements compared to manual measurement
range from 6% to 25%.
Description
Keywords
Detecting, tracking and extracting , pedestrian movements from video