TL;DR
This paper introduces PSD₁, a novel learning-based approach for detecting parking slots using surround-view images. To address the lack of standardized benchmarks, they also release a large-scale, manually labeled dataset of parking scenes, enabling more robust development and comparison of parking-slot detection algorithms.
What Problem Does This Paper Address?
The increasing demand for autonomous driving has spurred interest in automatic parking assistant systems (PAS). A critical component of PAS is accurately identifying vacant parking spaces. While various methods exist, they often fall into two categories: infrastructure-based (requiring external sensors and maps) and in-vehicle sensor-based. The latter, which relies on cameras, is further divided into free-space-based (detecting gaps between cars) and parking-slot-marking-based approaches. The paper focuses on the latter, highlighting two major challenges:
- Difficulty of Vision-Based Detection: Existing methods heavily rely on low-level visual features like lines and corners. These features are often unstable and indistinguishable under varying lighting conditions, background textures, and environmental noise, limiting detection accuracy.
- Lack of a Standard Benchmark: There's no publicly available, large-scale, labeled dataset for training and evaluating parking-slot detection algorithms. This hinders fair comparison and development of new methods.
Methods
The paper proposes a two-phase learning-based approach called PSDL (Parking-Slot Detection based on Learning):
- Marking-Point Detection:
- Feature Extraction: Instead of low-level features, PSD₁ focuses on "marking-points," which are the intersections of parking lines. These are considered more stable and distinguishable. Features used include normalized intensity, gradient magnitude, and oriented gradient magnitudes.
- Classifier: An AdaBoost framework with shallow decision trees is employed as a binary classifier. This classifier is trained to distinguish between marking-point patterns and non-marking-point patterns.
- Handling Rotations: To account for the varied orientations of marking-point patterns, multiple detectors are trained, each responsible for a specific angular range. This is achieved by rotating positive training samples.
- Color to Grayscale: For robustness against varying lighting, color images are converted to grayscale before processing.
- Parking-Slot Inference:
- Entrance-Line Candidate Generation: Pairs of detected marking-points are considered as potential entrance-line candidates. These candidates are validated by checking if the surrounding image patterns conform to predefined parking-slot models (T-shaped or L-shaped).
- Gaussian Line Templates: Six Gaussian line templates are used to examine the local image patterns around the marking-points. Simple rules are applied to these template responses to determine if a line segment is a valid entrance-line and to infer the parking-slot's orientation (clockwise or anti-clockwise).
- Contradiction Removal: A post-processing step removes conflicting entrance-line candidates to ensure consistency.
The system is designed to detect both perpendicular and parallel parking slots and works in both indoor and outdoor environments.
Datasets
- Newly Introduced Benchmark Dataset: The authors have created and made publicly available a large-scale dataset comprising 8600 surround-view images. These images were collected from typical indoor and outdoor parking sites using their experimental car.
- Training Set: Includes 5100 images for positive marking-point patterns (with center and orientation labeled) and 2400 images for negative samples.
- Testing Set: Consists of 500 labeled surround-view images for evaluating both marking-point detection and final parking-slot detection accuracy.
- Spatial Resolution: Each image has a resolution of 600x600 pixels, corresponding to a 9600mm x 9600mm physical area.
Key Insights and Ideas
- Marking-Points as Robust Features: The core insight is that using "marking-points" (intersections of parking lines) as features is significantly more robust and distinguishable than relying on primitive features like lines or corners, especially under challenging environmental conditions.
- Learning-Based Approach for Stability: Employing a data-driven learning strategy (AdaBoost) makes the detection process more resilient to variations in lighting and imaging conditions compared to traditional low-level vision algorithms.
- The Importance of a Standard Benchmark: The paper strongly advocates for and provides a benchmark dataset, recognizing its crucial role in advancing research by enabling fair comparisons and reproducible results.
- Multi-Detector Strategy for Orientation: The approach of training multiple detectors, each specialized for a range of orientations, effectively handles the directional variability of parking-slot markings.
Key Results
- Marking-Point Detection Performance: PSD₁ achieved a Log-Average Miss Rate (LAMR) of 18.82% on their benchmark dataset, outperforming other object detection methods like Roerei (23.77%) and significantly better than VJ, HoG+SVM, etc.
- Parking-Slot Detection Performance: When compared to state-of-the-art methods (Jung et al., Wang et al., Hamada et al., Suhr&Jung) at a high precision rate (around 98.7%), PSD₁ achieved a significantly higher recall rate of 92.38% (Table 1).
- Processing Speed: The system, implemented in C++ on an in-vehicle industrial computer, can process 20-25 surround-view image frames per second, meeting the requirements for most automatic PASs.
Limitations and Future Work
- Limitations:
- The paper acknowledges that the "depth" of the parking slot is determined by prior knowledge, implying it's not directly inferred from the visual data.
- When multiple parking slots are detected, the decision-making module is responsible for choosing the most appropriate one, suggesting this selection logic is external to the detection algorithm itself.
- Future Work: The authors suggest that PSD₁ can serve as a baseline for researchers developing more sophisticated parking-slot detection methods. The availability of their dataset and source code also facilitates future research in this area.
Summary updated