ROS2 UAV AI Applications

summarising UAV-related knowledge and AI-based UAV applications to get ideas for research topics.

UAV background

classification:
- fixed-wing, single rotor, fixed-wing hybrid, multirotor
characteristics:
- speed and flight time, payload, sensing equipment, software, range and altitude, controllers
applications:
- search and rescue
- infrastructure and construction inspection
- emergency medical services
- real-time monitoring of road traffic
- precision agriculture
- …
challenges:
- Efficient target detection
- unstable networks for swarm of drones
- resource allocation/energy optimisation
- limited transmission range
- standardisation and operational regulations
- security of sensitive data (pos, loc…)
- …
AI research directions:
- deep-learning-based novel strategies for UAVs, particularly for SAR mission

ai-uav

ai-uav-2

Detection - machine learning:
- deep learning algorithms for target detection
Navigation - automatic planning:
- optimal path planning in flight control system
- autonomous obstacle avoidance
Communication - distributed AI:
- coordinate intelligent behaviors among a group of autonomous mobile agents
- multi-UAV cooperative formation (UAV swarm)

Automated Wheat Disease Detection using a ROS-based Autonomous Guided UAV

image-based deep learning approach
- object detection model: YOLO V4, EfficientDet, MobileNet V3, VGG 16, Inception, ResNet 50, and EfficientNet-B0.
- CNN: weight-sharing ability
a custom dataset has been created and labeled
- data augmentation techniques: increase the number of training data
efficient mapping and navigation system is presented using a Gazebo simulation
- SLAM (simultaneous localization and mapping):
  - map: contains the static obstacles of the robot’s workspace
  - localization: enables the robot to map its environment while simultaneously estimating its own position regarding this map.
  - using sensors (camera, laser range-finder, ultrasonic sensor) to evaluate its distance to the nearby obstacles.
  - implemented and tested in a simulation environment
A 2D simultaneous localization and mapping algorithm is used for mapping the workspace autonomously.
Summary: ml for image classification, slam for exploration, A* algorithm for determining move trajectory.

uav-navi-ai

Real-time ROS Implementation of Conventional Feature-based and Deep-learning-based Monocular Visual Odometry for UAV

localization is one of the most important tasks for UAVs
- conventional feature-based methods:ORB-SLAM3
  - detection of key features in each image and match them on consecutive images to estimate the camera motions
- Deep-learning: SC-SfMLearner (self-supervised learning)
  - achieve high results on public dataset, but lack of real-time implementation in ROS for navigation system.
photo-realistic simulator, Flightmare, is used to test the implementation
- fast in collecting and computing a large number of images,
- but also integrate an accurate UAV dynamic
Visual Odometry (VO) benchmark dataset: KITTI, EuRoC, and CityScapes
evaluation process is done by integrating the aforementioned pose estimators into a real-time navigation system, benchmarking on streams of photo-realistic synthesized RGB data acquired from Flightmare.
- the outperforming of the features-based ORB-SLAM3 over learning-based counterpart, SC-SfMLearner
- learning-based algorithms are vulnerable to failure cases in which the drone trajectory contains excessive yaw rotations.

Twitter Facebook LinkedIn