v
ABSTRACT
Computer vision has become an increasingly popular area of
research. Computer vision always deals with providing non invasive solution
to the circumstances. In surveillance based computer vision application,
object detection and tracking are the basic steps of segmenting and tracking
an object of interest from a sequence of video frames. When the camera is
fixed, with consistent background object detection can be done with simpler
techniques. But, in aerial surveillance, the target along with background and
video camera are in motion. Also the size and shape of the target varies with
the altitude of flight. The proposed work is designed to detect and track the
targets in the video data captured by Unmanned Aerial Vehicle (UAV). An
UAV is an aircraft without pilot onboard. UAV DHAKSHA is a remote
operated unmanned miniature helicopter. It is designed to carry payload like
visual and acoustic sensors. It is capable of communicating with the
controller and returns payload data like video, acoustic signals, position,
speed, heading and altitude. The device is optimized for maximum efficiency,
low noise, less weight and higher endurance. Target detection and tracking
are the two approaches discussed in this thesis. A dataset of aerial videos with
varying circumstances and properties is collected by deploying the UAV in
real time environment and fed as input data to the designed algorithms.
vi
Target Detection in video sequence is the first step in any of the
surveillance applications. The designed target detection algorithm should be
accurate in the presence of varying lighting conditions, dynamic clutter and
variations in scene. Background subtraction is the most widely used technique
for detection of moving objects in the video. Most of the Background
subtraction techniques make a hypothesis that the content in the input video
sequence possess fixed background in which the object is moving. But, UAV
captures the video dynamically. Hence, Running Gaussian background
subtraction technique with adaptive setup to update the background frame by
frame is proposed and implemented. Also the performance of target detection
is improved by defining a fixed value to the tuning factor. A suitable value for
the tuning factor (α) is optimized statistically. The technique is compared with
few other existing techniques like Temporal frame differencing (TMF),
Running average background subtraction (RABS) and Temporal median
filtering (TMF) techniques.
Target tracking is the process of estimating the timely location of
one or multiple targets of interest in a video stream captured by a visual
camera. Mean shift tracking is a non-parametric, kernel based tracking
technique which provides accurate localization and matching of target without
expensive search. It works on iterative passion. It computes the mean value
for the position of pixel in the current window and shifts to new position in
the next frame which possesses the mean value and continues till it satisfies
the conditions. However, the traditional mean shift tracking approach works
on assumption that the target differs significantly from background. But, in
vii
cases like aerial videos it is difficult to discriminate the background and the
target. Thus an adaptive background model based Mean Shift tracking
technique is proposed and implemented. Here the background model is
defined and updated frame by frame thereby the problem of occlusion gets
eliminated. The same video used for target detection is fed to the target
tracking algorithm. The algorithm is compared with Traditional mean shift
tracking technique (TMST) and Continuously adaptive mean shift technique
(CAM Shift) to evaluate and justify the efficiency of the proposed algorithm.
In this work, some novel approaches to object detection and tracking are
done. The works are simulated using MATLAB and the performance is
evaluated by image based and video based metrics.