Autonomous Model Selection for Surface Classification via Unmanned Aerial Vehicle
Poster Number
8
Introduction/Abstract
Gathering data in remote or sensitive areas can provide numerous obstacles to researchers. Deploying and retrieving sensors can be difficult or hazardous, such as when trying to monitor protected wetlands. This requires a system that can perform this task with as little impact on the environment as possible, and with minimal risk to the researchers.
Purpose
The purpose of this research is to create a system that can deploy sensor networks autonomously with minimal impact on the environment. Part of this process requires a UAV to know what type of ground it is trying to install a sensor into, which is the focus of this work.
Method
An effective method for accomplishing this task is to use a quadcopter with a gripper attachment. A motion capture system tracks the location of the quadcopter and allows precise automated control via a feedback loop. The UAV has an accelerometer attached to it that records the impact and rebound during landing. This data is collected autonomously and sent to a central system. This system trains many machine learning classifiers, evaluates them with user-defined settings, and saves the best one. The classifier is loaded onto the UAV so it can classify new surfaces based on the landings.
Results
The meta classification system generates a large number of models and selects the one with the best accuracy and fewest features. Two test data sets were collected, one labeling samples as hard or soft, and the other labeling samples one of four surface types. Each set contains 11 features. For the four surface data, our system resulted in a model with 75% accuracy while using 3 features, making it an efficient choice. This beats both manually trained classifiers (63% with 11 features) and the classifier selected by AutoWEKA (70% with 11 features). We see similar results with the hard/soft data, with our system providing high accuracy models using as few as 2 features.
Significance
The meta classification algorithm provides automated generation of high accuracy efficient classifiers suitable for robotics applications where efficiency is key. While we use it with time series data, it can be used for any classification task.
Location
DUC Ballroom A&B
Format
Poster Presentation
Poster Session
Afternoon
Autonomous Model Selection for Surface Classification via Unmanned Aerial Vehicle
DUC Ballroom A&B
Gathering data in remote or sensitive areas can provide numerous obstacles to researchers. Deploying and retrieving sensors can be difficult or hazardous, such as when trying to monitor protected wetlands. This requires a system that can perform this task with as little impact on the environment as possible, and with minimal risk to the researchers.
