Deep-Learning-Based Parking Area and Collision Risk Area Detection Using AVM in Autonomous Parking Situation

Area of Science:

• Computer Vision
• Deep Learning
• Autonomous Driving Systems

Background:

• Parking situations present complex challenges for vehicle safety and efficiency.
• Accurate detection of parking areas and potential collision risks is crucial for advanced driver-assistance systems (ADAS).
• Existing methods often struggle with real-time performance and comprehensive area identification.

Purpose of the Study:

• To propose a novel bird eye view image detection method for simultaneously identifying parking areas and collision risk zones.
• To develop a deep learning architecture capable of real-time analysis of around view monitor (AVM) images.
• To evaluate the performance of the proposed method against conventional approaches.

Main Methods:

• Utilized deep learning algorithms, specifically semantic segmentation and area detection.
• Developed a main architecture based on a harmonic densely connected network (HarDNet) and a cross-stage partial network (CSPNet), termed attention CSPHarDNet.
• Trained the model on a dataset of around view monitor (AVM) images generated from four 190° wide-angle cameras at Chungbuk National University parking lot.

Main Results:

• The attention CSPHarDNet model achieved 81.89% mean Intersection over Union (mIoU) and 18.36 Frames Per Second (FPS) on a NVIDIA Xavier environment.
• Successfully visualized available parking areas by detecting parking lines, parking spaces, and drivable zones.
• Effectively visualized collision risk areas by identifying undetected regions in semantic segmentation.

Conclusions:

• The proposed attention CSPHarDNet model demonstrates real-time applicability for parking situations.
• The method offers superior performance compared to conventional HarDNet, enhancing parking safety and efficiency.
• This approach contributes to the advancement of intelligent parking systems and ADAS.