Target detection and classification via EfficientDet and CNN over unmanned aerial vehicles

Area of Science:

• Computer Vision
• Artificial Intelligence
• Traffic Engineering

Background:

• Conventional vehicle detection methods struggle with computational demands and data variability.
• Aerial imagery presents unique challenges for accurate vehicle analysis.
• Existing systems require adaptation for diverse data collection.

Purpose of the Study:

• To introduce an innovative technique for vehicle classification and recognition in aerial image sequences.
• To enhance the efficiency and accuracy of traffic monitoring systems.
• To overcome limitations of current vehicle detection approaches.

Main Methods:

• Image enhancement using noise reduction and Contrast Limited Adaptive Histogram Equalization (CLAHE).
• Object identification via contour-based and Fuzzy C-means segmentation (FCM).
• Vehicle detection/identification with EfficientDet; feature extraction using Accelerated KAZE (AKAZE), Oriented FAST and Rotated BRIEF (ORB), and Scale Invariant Feature Transform (SIFT).
• Classification using Convolutional Neural Network (CNN) and ResNet.

Main Results:

• The proposed model achieved 96.6% accuracy on the Unmanned Aerial Vehicle Intruder Dataset (UAVID).
• Achieved 97% accuracy on the Vehicle Aerial Imagery from a Drone (VAID) dataset.
• Demonstrated superior performance compared to existing methods.

Conclusions:

• The developed model significantly improves vehicle detection and classification in aerial images.
• Offers notable advancements for intelligent traffic monitoring systems.
• Surpasses current methodologies in accuracy and adaptability.