Related Experiment Video
Updated: May 16, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
438
Analysis of Deep Learning Techniques for Vehicle Detection and Reidentification Using Data from Multiple Drones and
Felipe P A Euphrásio1,2, Rafael M DE Andrade2,3, Elcio H Shiguemori2,3
1Instituto Tecnológico de Aeronáutica - ITA, Praça Marechal Eduardo Gomes, 50, Vila das Acácias, 12228-900 São José dos Campos, SP, Brazil.
Anais Da Academia Brasileira De Ciencias
|April 2, 2025
Summary
This study enhances vehicle detection and re-identification using multiple convolutional neural networks (CNNs) and tracking algorithms for drone surveillance. The methods achieved up to 91% accuracy, improving vehicle tracking in complex environments.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Robotics
Background:
- Vehicle detection and re-identification in dynamic environments like drone-monitored highways is challenging due to image variability.
- Existing methods require improvement to handle diverse capture angles and conditions from multiple drones.
Purpose of the Study:
- To propose and evaluate a robust solution for vehicle detection and re-identification using various convolutional neural networks (CNNs).
- To assess the generalization capabilities of CNNs with varied drone imagery for improved accuracy.
Main Methods:
- Combined YOLOv4 for detection and DeepSORT for tracking.
- Integrated and evaluated CNN models: VGG16, VGG19, ResNet50, InceptionV3, and EfficientNetV2L.
- Utilized a public dataset and a custom dataset from a drone swarm for evaluation.
Main Results:
- ResNet50 achieved 55% average accuracy in the first experiment.
- VGG19 demonstrated the highest accuracy at 91% in the second experiment.
- The proposed methods successfully distinguished different vehicle models and adapted to drone-captured data.
Conclusions:
- The integrated approach effectively addresses challenges in drone-based vehicle surveillance.
- CNNs, particularly VGG19, show strong performance in vehicle re-identification from diverse drone perspectives.
- The study validates the adaptability and accuracy of the proposed methods for real-world applications.
