Vision-based autonomous object tracking using multi-rotor UAV

Abstract

In this paper- we are presenting a real-time method to detect and track an autonomous object
using visual processing on an unmanned aerial vehicle using an on-board companion
computer (Jetson-TX1) for image processing. The profile of objects, frame rate of images,
and unexpected motion make it hard to detect and track the object for a long period. To cater
to this, we came up with an algorithm that was developed for long-term tracking which makes
use of Discriminative correlation filter with Channel and Spatial Reliability. The major
restriction of our algorithm arises in the presence of occlusion, which was solved by creating
a region of interest in the center of the frame in which the object will always reside. If the
object exits the center region a command of left, right, top or bottom will be generated
following the position of the object relative to our center position. These commands will be
communicated to the UAV via Mavlink protocol. Experimental results show that we have
achieved a long period of tracking with a good frame rate and eliminating spurious events and
misdetections.