Institutional Repository of Vilnius University
Drone detection using state space models /
Abstract
dc:descriptionIn recent years the accessibility of unmanned aerial vehicles, also known as drones, has drastically increased. With increased accessibility new challenges related to misuse have emerged: smuggling, espionage, and disruptions of airport operations. To address this issue, accurate and fast automated drone detection systems are required. Most modern object detection models rely on convolutional neural networks or transformer-based architectures. Although transformers provide a strong global context understanding, their quadratic time complexity limits their use in real-time applications. State space models (SSMs) have recently emerged as an alternative to transformers with their linear time complexity and efficiency in long-range dependency modelling. This study investigates the application of SSM-based vision architectures to real-time drone detection tasks by integrating VMamba modules into the YOLO26 architecture. Several hybrid YOLO26-VMamba models were created by inserting vision selective scan (VSS) blocks into different parts of the YOLO26 model, including the backbone, neck and as a replacement for the attention module. Additionally, the recently released Mamba3 kernel was integrated into the VSS module and compared to Mamba and Mamba2 versions. Training and testing were done on COCO and drone detection datasets. The results show that VMamba-based hybrid models can be successfully adopted for real-time drone detection tasks; however, their effectiveness is highly dependent on the placement of VSS blocks within the architecture and the Mamba kernel variant used. Some hybrid models achieved a slight improvement in detection accuracy at the cost of inference speed compared to the baseline YOLO26 model.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Balčiūnas, Justas,
Subjects
dc:subject × 1Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- eng
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD308332287&prefLang=en_US