Development of a machine-learning-based detection and tracking algorithm for drones
The proliferation of drones around the world was sparked by the reduction in the cost of owning a drone. As such, numerous drone research, worldwide, began to sprout. Majority of these researches focused on the applications of drones using computer vision as well as artificial intelligence. Many res...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Final Year Project |
Language: | English |
Published: |
Nanyang Technological University
2021
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/150863 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | The proliferation of drones around the world was sparked by the reduction in the cost of owning a drone. As such, numerous drone research, worldwide, began to sprout. Majority of these researches focused on the applications of drones using computer vision as well as artificial intelligence. Many researchers developed detections and tracking algorithms that focused on bounding boxes. Some researchers dedicated their work to semantic segmentation using aerial imagery or street views. Few had tried integrating detections and tracking for drones. However, fewer had tried using semantic segmentation for detection and tracking using drones [1], [2]. It is a challenge to achieve real time semantic segmentation for detection and tracking on drones as the drones had a limited amount of computation power. Since edge devices, often, did not have high processing power [3], steps must be taken to ensure that the algorithm was efficient in accommodating edge devices. After comparing several algorithms, DeepLabv3+ with Mobilenetv2 backbone was the fastest algorithm and was tested using one of the best edge devices available, NVIDIA Jetson TX2, for detections and tracking. Drones were able to conduct real time detections and tracking operations. However, the frame rate of this operation was lower than expected due to hardware limitations. This project indicated that efficient semantic segmentation detections and tracking using drones would be possible with future improvements in the processing power of hardware and/or improvement in the efficiency of algorithms. |
---|