High-Dynamic-Range Tone Mapping in Intelligent Automotive Systems.

Area of Science:

• Computer Vision
• Automotive Engineering
• Image Processing

Background:

• Intelligent driver assistance systems rely on accurate vulnerable road user (VRU) detection.
• Standard imaging sensors struggle in high-contrast lighting, limiting perception system performance.
• High-dynamic-range (HDR) imaging offers a solution but requires tone mapping for compatibility with existing detectors.

Purpose of the Study:

• To evaluate the impact of tone mapping on object detection performance for HDR images.
• To develop and optimize an HDR tone mapping method that preserves natural image appearance while enhancing object detection.
• To introduce a novel detection-informed tone mapping (DI-TM) approach for vehicle perception systems.

Main Methods:

• Utilized a lightweight convolutional neural network (CNN) for tone mapping HDR video frames to standard 8-bit representation.
• Developed a novel training approach, detection-informed tone mapping (DI-TM), integrating object detection feedback.
• Evaluated DI-TM's effectiveness, robustness, and performance against a state-of-the-art tone mapping method across various scene conditions.

Main Results:

• The proposed DI-TM method significantly improved detection performance metrics in challenging dynamic range conditions.
• DI-TM enhanced the detection F2 score by 13% in challenging conditions compared to other methods.
• A substantial 49% increase in F2 score was observed when comparing DI-TM with standard dynamic range (SDR) images.

Conclusions:

• Detection-informed tone mapping (DI-TM) is effective for optimizing HDR imaging in vehicle perception.
• The DI-TM method provides superior object detection performance in challenging lighting scenarios.
• This approach enhances the safety and reliability of intelligent driver assistance systems.