OBJECT DETECTION METHOD TOWARDS AUTONOMOUS INSULATOR INSPECTION WITH UNMANNED AERIAL VEHICLE
This undergraduate thesis explores the usage of single-stage CNN models to detect insulators found in aerial images and measures their applicability in UAV onboard systems. The project is motivated by existing methods in literature which unfortunately does not perform well in said environments...
Saved in:
| Main Author: | |
|---|---|
| Format: | Final Project |
| Language: | Indonesia |
| Online Access: | https://digilib.itb.ac.id/gdl/view/55932 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Institut Teknologi Bandung |
| Language: | Indonesia |
| Summary: | This undergraduate thesis explores the usage of single-stage CNN models to detect insulators
found in aerial images and measures their applicability in UAV onboard systems. The project is
motivated by existing methods in literature which unfortunately does not perform well in said
environments. The proposed methods include modifying a baseline network of YOLOv2 with SPP
(spatial pyramid pooling) blocks and optimizing its bounding box regression function. Methods
that involve limiting the filter depth within each convolution layer are also reviewed. It is
concluded that the usage of SPP blocks and CIoU loss increases the overall network performance
without sacrificing inference speed. However, networks with limited filter depth are much more
suitable for onboard usage. One of such design is SF-YOLO, with computation cost of 3,842
BFLOP (29% lower than YOLOv3-tiny, 86% lower than proposed baseline) while retaining AP50
score higher than 0.9, and thus can be further used for autonomous navigation subsystems due to
its ability to run at > 30FPS with proper edge devices. |
|---|