Stereo Camera-Based Aerial Target Geolocation in a Ground-Based Platform Environment
Across diverse surveillance, rescue, and scientific research applications, there is a specific need to perform on-the-ground target object detection and localization from highly efficient aerial platforms. The aerial target geolocation task has seen rapid development in recent years, contributed...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | en_US |
| Subjects: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Across diverse surveillance, rescue, and scientific research applications, there is a
specific need to perform on-the-ground target object detection and localization from
highly efficient aerial platforms. The aerial target geolocation task has seen rapid
development in recent years, contributed by the explosive growth of unmanned aircraft
vehicles (UAVs) and artificial intelligence (AI) computing. The existing solutions of
using the monocular camera as the vision-based geolocation sensor exhibited a high
chance of geolocation error and inconvenience requirement of pre-operation calibration
of the camera's intrinsic parameter. Studies on the potential of using stereo vision
technology for geolocating targets from aerial platforms are still scarce, especially when
integrated with automated detection and edge processing. This study aimed to develop
a stereo vision-based aerial target geolocation system with excellent effectiveness and
operational flexibility. This research explores three modules: (1) stereo vision-based
geolocation modeling, (2) deep learning-based object detection modeling, and
(3) evaluation of edge computer performance. The novel stereo vision-based aerial
target geolocation algorithm is formulated by using a constructive model development
technique, whereby the stereo vision point cloud information of the detected target is
used as the distance and angle parameters in the radar target tracking model and
projected coordinate system. The relevant Darknet-based object detection models are
evaluated, each trained with the VisDrone2018 aerial training dataset. The detection
and geolocation module has been executed in the Nvidia Jetson TX2 edge computing
platform to determine the system's feasibility. The evaluation and validation of the
proposed geolocation, detection, and processing modules are performed using the
ground-based platform field test experiments and MATLAB. The field experiment
results validated the proposed stereo vision-based aerial target geolocation,
demonstrating the geolocation accuracy of 0.53-meter mean error at the 5-meter testing
height. Besides, the selected detection module, the YOLOv4 model, scored a detection
accuracy of 30.69% mean Average Precision (mAP) value when tested on the official
test tool, and its detection speed satisfied the safety requirements of 2 frames per second
(FPS) for aerial applications. Further, the edge computer platform exhibited minimal
power consumption of 5 Watts and maintained the manufacturer's system operating
standards. These findings demonstrate the feasibility of the proposed stereo vision based
aerial target geolocation system to be used as a system payload for aerial
platforms and serve as a reference framework for Search and Rescue (SAR) agencies in
detecting distressed humans. |
|---|