1 results listed
The rapid advancement in Unmanned Aerial Vehicle
(UAV) technology in recent years has led to their widespread
adoption across various sectors. However, positioning challenges
encountered by UAVs in challenging environments such as
indoor spaces, deep canyons, and military operation zones have
emerged as a significant research concern. Given the critical
importance of precise positioning information for safe UAV
operations, the development of alternative solutions has become
imperative in situations where satellite-based positioning systems
prove inadequate. In this context, extensive research has been
conducted on visual, inertial, and visual-inertial fusion
approaches in the literature. Recent research in this field has
particularly focused on deep learning-based methods, which have
demonstrated effective performance even in the presence of
complex environmental conditions and noisy inertial data. In
existing studies, the sequential order has been preserved during
training processes, considering the time series nature of the data.
However, in fusion-based approaches, the potential of
Convolutional Neural Network (CNN) architectures to operate
independently of time series has not been adequately
investigated. This research proposes a novel model that combines
CNN-based visual feature extraction with Bidirectional Long
Short-Term Memory (BiLSTM)-based inertial feature extraction.
The original contribution of this study lies in its systematic
examination of the effects of shuffling operations on the dataset.
Experimental results reveal that despite the time series nature of
the data in visual-inertial fusion models, the shuffling operation
leads to significant improvements in model performance.
International Conference on Advanced Technologies, Computer Engineering and Science
ICATCES
Mahmut Karaaslan
Ersin Kaya