Convolutional transform learning based fusion framework for scale invariant long term target detection and tracking...

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unmanned aerial vehicles (UAVs) are increasingly used for environmental monitoring and computer vision (CV) tasks due to their mobility.
• Conventional UAV detection models struggle with scale variations and complex environments, leading to missed detections and false alarms.
• Accurate target detection and tracking are crucial for UAV applications in military, urban planning, and wildlife monitoring.

Purpose of the Study:

• To develop a novel long-term target detection and tracking model for UAVs.
• To enhance the robustness and accuracy of UAV-based computer vision systems in scale-variant environments.
• To address challenges in UAV identification caused by camera motion and environmental complexity.

Main Methods:

• The proposed LTTDT-UAVDFCTL model utilizes image pre-processing with a median median-enhanced wiener filter (MEWF) for noise reduction.
• Object detection is performed using YOLOv8, with feature extraction via a deep fusion backbone (VGG16, CapsNet, EfficientNetB7).
• Graph convolutional neural networks (GCN) are employed for tracking, optimized by a hybrid sine cosine whale optimization algorithm (SCWOA).

Main Results:

• The LTTDT-UAVDFCTL model demonstrated superior performance in target detection and tracking on the VisDrone dataset.
• Achieved a mean Average Precision (mAP) of 80.13%, surpassing existing models.
• The model effectively handles scale variations and improves accuracy while reducing false positives.

Conclusions:

• The developed LTTDT-UAVDFCTL model offers a robust solution for long-term target detection and tracking with UAVs.
• The deep fusion approach significantly enhances performance in challenging, scale-variant environments.
• This research contributes to advancing UAV-based computer vision applications through improved detection and tracking capabilities.