Improving the Automatic Counting in a Region of Interest in Videos of Urban Traffic taken by Drones using Video Stabilization
Abstract
Due to the rapid progress of video analytics,
traffic monitoring has emerged as a vital method for
collecting information about traffic conditions. This
paper presents an improved, automated system for
counting vehicles and pedestrians from drone footage.
The system allows a user to define any counting line
within the video scene. Our four-module pipeline
consists of: (1) a robust video stabilization module that
compensates possible movements of the drone, (2) a
deep learning-based object detection model trained on
a custom dataset recorded in Yucat´ an, Mexico to identify
eight distinct classes (car, truck, bus, van, motorcycle,
bike, moto-taxi, and pedestrian), (3) an object tracking
algorithm that analyzes detection bounding boxes, and
(4) a counting module that tallies objects crossing the
user-defined segment. The proposed counting system
extends our previous work that did not use a stabilization
module [6].
traffic monitoring has emerged as a vital method for
collecting information about traffic conditions. This
paper presents an improved, automated system for
counting vehicles and pedestrians from drone footage.
The system allows a user to define any counting line
within the video scene. Our four-module pipeline
consists of: (1) a robust video stabilization module that
compensates possible movements of the drone, (2) a
deep learning-based object detection model trained on
a custom dataset recorded in Yucat´ an, Mexico to identify
eight distinct classes (car, truck, bus, van, motorcycle,
bike, moto-taxi, and pedestrian), (3) an object tracking
algorithm that analyzes detection bounding boxes, and
(4) a counting module that tallies objects crossing the
user-defined segment. The proposed counting system
extends our previous work that did not use a stabilization
module [6].
Keywords
Automatic counting, video stabilization, object detection, object tracking, deep learning