Loading...
Publication
ADVANCED UAV MONITORING: DEEP LEARNING FOR MULTI -TARGET DETECTION, TRACKING, AND WILDFIRE PREDICTION
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 37.86 MB | Adobe PDF |
Abstract(s)
Rising global fire incidents necessitate effective solutions, making forest surveillance
crucial. Current methods require substantial investment and labor but are often
ineffective. This work proposes a comprehensive monitoring solution utilizing Unmanned
Aerial Vehicles (UAVs) to integrate visible and infrared images for real-time
detection of people, vehicles, and fires, addressing limitations in low-light conditions,
fog, or smoke.
We propose a new system architecture for real-time UAV footage transmission,
processing, and analysis on a cloud server. For the detection of people and vehicles,
we propose a new 4-channel object detection model that significantly improves
precision metrics compared to traditional state-of-the-art models that utilize only
RGB images. Additionally, our model performs better in conditions unfavorable
to RGB images, successfully identifying objects in low light and reduced visibility.
To train our model, we present a labeled dataset with aligned thermal and visible
images from an aerial perspective.
In order to enable object tracking in our solution, which refers to detecting
and maintaining a unique identifier for each detection, we propose SAME, a new
approach to Multiple Object Tracking (MOT) re-identification. The proposed
model is designed to extend the capabilities of existing detectors by using the
high-dimensional features they extract as inputs to a transformer-based architecture.
This method applies attention and transformers to measure the similarity between
tracks across multiple frames, significantly improving re-identification performance.
SAME employs transformers to enable past context retrieval, standing out for its
modularity while achieving competitive results in known datasets such as MOT17
and BDD100K.
Finally, we introduce FireSeq, a novel approach leveraging state-of-the-art deep
learning techniques such as VQ-VAEs and Transformers to model wildfire progression
in real time. To support this research, we developed the FireSeq dataset, which
includes both RGB and infrared (IR) aligned imagery capturing the behavior of
wildfires from an aerial perspective. Additionally, the FireSeq dataset includes a
labeled multi-class subset designed for early wildfire detection. FireSeq demonstrates
a high degree of accuracy in predicting future frames of wildfire footage. The three developed components represent innovative research approaches that
together form a comprehensive and robust wildfire monitoring solution. This marks
a significant advancement in wildfire prevention and proactive management. By
enabling continuous real-time monitoring, detection, and tracking, our solution
supports critical applications such as risk analysis, crowd management, and searchand-
rescue operations. Furthermore, it introduces a novel method for predicting
and detecting wildfire progression, aimed at enhancing early detection capabilities
and improving mission planning efficiency.
Description
Keywords
Recolha de dados Processamento de dados Tratamento de imagem UAV’s Modelo FireSeq Incêndio florestal- deteção Computação móvel Veiculo aéreo não tripulado