A Resource-Efficient CNN-Based Method for Moving Vehicle Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Traffic Engineering

Background:

• Convolutional Neural Networks (CNNs) offer high accuracy for Automated Vehicular Surveillance (AVS) but are computationally intensive.
• Background Subtraction (BS) methods are lightweight but lack the detail for complex traffic analysis.
• A need exists for efficient AVS systems that balance accuracy and computational cost.

Purpose of the Study:

• To propose a hybrid framework combining BS and CNNs for efficient AVS.
• To reduce the computational complexity of CNN-based AVS systems.
• To maintain or improve detection accuracy while decreasing computational load.

Main Methods:

• A Background Subtraction (BS) module preprocesses video feeds to isolate moving objects, creating image candidates.
• A Convolutional Neural Network (CNN) detector is applied to these candidates, optimizing convolution operations.
• Four state-of-the-art CNN architectures were evaluated as base detectors.

Main Results:

• The proposed framework achieved significant computational complexity reduction across various CNN models.
• Complexity reduction increased with CNN model size, e.g., 30.6% for YOLOv5s and 52.2% for YOLOv5x.
• Detection accuracy was maintained despite the reduced computational load.

Conclusions:

• The integrated BS-CNN framework offers an effective solution for computationally efficient AVS.
• This approach enhances the feasibility of deploying advanced AVS systems in resource-constrained environments.
• The method successfully addresses the trade-off between accuracy and computational cost in vehicular surveillance.