A Study on Data Selection for Object Detection in Various Lighting Conditions for Autonomous Vehicles.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Automotive Engineering

Background:

• Advanced Driver Assistance Systems (ADAS) and autonomous vehicle technology are rapidly evolving.
• Automated object detection is critical for autonomous driving but struggles in low-light conditions.
• Poor visibility of objects in low light significantly impacts the performance of detection algorithms.

Purpose of the Study:

• To investigate how the composition of training data affects object detection performance in low-light conditions.
• To evaluate the impact of varied outdoor scene lighting (time of day) on deep neural network performance.
• To identify challenges in training neural networks for robust object detection.

Main Methods:

• Experiments were conducted using a popular public database and common object detection architectures.
• The study analyzed the effect of different combinations of outdoor images captured at various times of day.
• Deep neural networks were trained and evaluated based on training data composition.

Main Results:

• An appropriate balance of object classes and illumination levels in training data leads to more robust performance.
• Including images from dusk and dawn conditions significantly improves object detection.
• Performance gains were observed for both day and night low-light scenarios.

Conclusions:

• Training data composition is a key factor in achieving reliable object detection for autonomous vehicles.
• Strategic inclusion of varied lighting conditions, particularly twilight, enhances algorithm robustness.
• Further research into data augmentation for low-light conditions is warranted.