Adopting the YOLOv4 Architecture for Low-Latency Multispectral Pedestrian Detection in Autonomous Driving

Area of Science:

• Computer Vision
• Autonomous Systems
• Machine Learning

Background:

• Pedestrian detection is crucial for autonomous driving safety, requiring minimal latency.
• Multispectral imaging (RGB + thermal) enhances robustness across diverse conditions.
• Existing deep learning models achieve high accuracy but lack evaluation in real-time obstacle avoidance scenarios.

Purpose of the Study:

• To evaluate the YOLOv4 neural network for multispectral pedestrian detection.
• To assess its feasibility for low-latency obstacle avoidance in autonomous vehicles.
• To identify the optimal YOLOv4 variant and fusion strategy for automotive applications.

Main Methods:

• Adaptation of the You Only Look Once version 4 (YOLOv4) real-time neural network detector.
• Implementation of a middle fusion approach within the YOLOv4 Tiny architecture.
• Evaluation using the KAIST multispectral pedestrian dataset.

Main Results:

• YOLOv4 demonstrates comparable accuracy to state-of-the-art methods for multispectral pedestrian detection.
• The detector achieves high computational efficiency, suitable for low-latency decision-making.
• The YOLOv4 Tiny variant with middle fusion provided the best accuracy-efficiency trade-off.

Conclusions:

• YOLOv4 is a viable architecture for real-time multispectral pedestrian detection in autonomous driving.
• The YOLOv4 Tiny middle fusion approach is particularly promising for safety-critical automotive applications.
• Further research can leverage this for improved vehicle perception and collision avoidance systems.