RPNet: Rotational pooling net for efficient Micro Aerial Vehicle trail navigation

• Published in: Engineering applications of artificial intelligence Vol. 116; p. 105468
• Main Authors: Agyemang, Isaac Osei, Zhang, Xiaoling, Acheampong, Daniel, Adjei-Mensah, Isaac, Gyamfi, Enoch Opanin, Arhin, Joseph Roger, Ayivi, Williams, Amuche, Chikwendu Ijeoma
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01-11-2022
• Subjects: Autonomous; Drone; Gabor filters; Micro Aerial Vehicle; Rotational pooling; Drone; Micro Aerial Vehicle; Gabor filters; Autonomous; Rotational pooling
• ISSN: 0952-1976; 1873-6769
• DOI: 10.1016/j.engappai.2022.105468

Smart agile Micro Aerial Vehicles are becoming ubiquitous in industrial applications; however, their computational efficacy poses a constraint limiting their deployment. The challenge mentioned above is mitigated in the context of a Micro Aerial Vehicle vision-based navigation task. RPNet, a computational-efficient model, is proposed as a Micro Aerial Vehicle navigational controller. RPNet comprises a sequential arrangement of generic imaginary and real Gabor filters for computation reduction. Further, a novel rotational pooling mechanism that induces online feature descriptors augmentation is proposed and plugged after each convolutional block in RPNet for robustness and increase in performance. RPNet is initially trained on synthetic data for domain knowledge and further trained and tested in a real-world setting using a Micro Aerial Vehicle. Extensive experimental verification of RPNet based on four evaluation metrics shows satisfactory performance compared to the reference trajectories. Further, via comparisons, RPNet attains better error distribution of about ±5 m, and computational conservation of around 9% than the first runner-up comparator among the state-of-the-art models in the vision-based Micro Aerial Vehicle navigation task.
•Rotational pooling feature map augmentation improves CNNs/DCNNs’ resilience and performance.•RPNet, a computationally efficient model for MAV autonomous navigation, is proposed.•RPNet’s engineering practicability is shown via a vision-based trail navigation task.•Comparisons indicate that RPNet reduces computation by about 9% during convolutions.•Error distribution of RPNet in MAV navigational task is about ±5m better than the comparators.