Deep Multimodal Detection in Reduced Visibility Using Thermal Depth Estimation for Autonomous Driving.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Systems

Background:

• Convolutional Neural Networks (CNNs) excel in object detection for autonomous driving but struggle in poor weather.
• Haze and fog significantly degrade the performance of existing detection systems, posing safety risks.

Purpose of the Study:

• To develop an effective image dehazing network for autonomous driving systems.
• To improve the robustness and accuracy of moving target detection in adverse weather conditions like fog.

Main Methods:

• Proposed an image dehazing network that estimates weather conditions and removes haze.
• Integrated the restored RGB images with thermal images.
• Utilized two You Only Look Once (YOLO) object detectors with late fusion for independent detection.

Main Results:

• The proposed dehazing network demonstrated superior performance compared to existing models.
• Restored images taken in foggy conditions to a normal state.
• Achieved up to a 22% improvement in detection accuracy in dense fog environments through RGB-thermal fusion.

Conclusions:

• The developed image dehazing network enhances object detection capabilities for autonomous driving in poor visibility.
• Fusion of restored RGB and thermal images offers a robust solution for safe autonomous navigation in challenging weather.