Pedestrian Detection Using Multispectral Images and a Deep Neural Network

Area of Science:

• Computer Vision
• Automotive Safety
• Artificial Intelligence

Background:

• Vehicle-pedestrian crashes are a major cause of fatalities and injuries.
• Current Advanced Driver-Assistance Systems (ADAS) and autonomous vehicle pedestrian detection systems struggle with performance reduction in low-light conditions.
• Reliable pedestrian detection is crucial regardless of ambient lighting.

Purpose of the Study:

• To evaluate pedestrian detection performance across various lighting conditions.
• To propose the use of multispectral imaging and optimized deep neural networks to enhance detection accuracy.
• To improve the reliability and efficiency of pedestrian detection systems.

Main Methods:

• Comparative analysis of pedestrian detection using RGB, thermal, and multispectral image formats.
• Development and optimization of deep neural network architectures for pedestrian detection.
• Evaluation of detection accuracy and processing time trade-offs.

Main Results:

• Multispectral images demonstrated superior performance for pedestrian detection in diverse lighting conditions compared to RGB and thermal.
• The proposed deep neural network achieved a 6.9% improvement in pedestrian detection accuracy over baseline methods.
• Optimizations reduced processing time by 22.76% with a minimal 2% decrease in detection accuracy.

Conclusions:

• Multispectral imaging is the optimal solution for robust pedestrian detection across varying lighting conditions.
• The developed deep neural network effectively enhances pedestrian detection accuracy and efficiency for ADAS.
• Balancing processing time and detection accuracy is feasible through network optimization.