Parking Slot Detection on Around-View Images Using DCNN

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Vision-based parking slot detection is challenging due to complex visual environments and incomplete around-view image data.
• Existing methods often rely on cumbersome, multi-step processes using pre-trained models.
• There is a need for a more robust, simple, and efficient parking slot detection system.

Purpose of the Study:

• To propose a novel, robust, and simple parking slot detection method using deep convolutional neural networks (DCNNs).
• To represent parking slots as directional entrance lines for simplified detection.
• To achieve high accuracy and real-time performance in diverse parking scenarios.

Main Methods:

• A deep convolutional neural network (DCNN) detector was designed for simultaneous regression and classification of directional entrance lines.
• Parking slots are represented as directional entrance lines, capturing type, position, length, and direction.
• The complete parking slot is inferred using DCNN detection results and prior geometric information.

Main Results:

• The proposed method achieved state-of-the-art performance on the ps2.0 dataset with 99.68% precision and 99.41% recall.
• Demonstrated satisfying generalization capabilities on a self-annotated parking slot dataset.
• Achieved a real-time detection speed of 13 ms per frame on a Titan Xp GPU.

Conclusions:

• Converting parking slots into directional entrance lines simplifies detection.
• The specially designed DCNN detector efficiently detects various parking slot types.
• This method offers a robust, simple, and fast solution for vision-based parking slot detection.