Abstract
The autonomous parking of vehicles requires the ability to accurately locate an available parking slot in the vicinity of a vehicle. Since parking slots have a variety of shapes and colors, may be occluded by obstacles, or look different due to surroundings such as lighting, accurately locating them can be a challenging task. In this paper, we propose a context-based parking slot detection method inspired by the process of a human driver finding a parking slot. Our method consists of two deep network modules: a parking context recognizer and parking slot detector . The parking context recognizer identifies the parking environment (type, angle, and availability of a parking slot), whereas the parking slot detector locates the exact position of a parking slot by multiple type-based fine-tuned detectors with rotated anchor boxes and a rotated non-maximal suppression. In addition, we release a realistic parking slot dataset, which comprises 22817 images of parking slots having various attributes and external conditions. We also propose a new evaluation metric for parking slot detection, reflecting whether a vehicle can be parked within the detected parking slot. Through comparison and ablation study in experiments, we demonstrate that our method outperformed the previous deep-learning-based methods, along with having a short operation time. The source codes and the dataset are available at https://github.com/dohoseok/context-based-parking-slot-detect/ .
Highlights
An autonomous parking system is essential for autonomous vehicles
We propose a performance evaluation metric that considers when a parking slot detection method is applied to an autonomous parking system
PARKING CONTEXT RECOGNIZER In this experiment, we evaluated the performance of the parking context recognizer (PCR) using various networks as the backbone
Summary
An autonomous parking system is essential for autonomous vehicles. Such a system must first detect the parking slot and control the vehicle to park it in the designated slot. Previous methods for vision-based parking slot detection focused on finding the accurate location of the parking slot directly from a surrounding image This is because most parkingassist systems installed in many mass-produced vehicles are operated after a person manually drives the car to a space where parking is available. This paper proposes a two-stage parking slot detection method based on context information of the entire image. The proposed two-stage approach reduces the amount of computation and false positive errors because the parking slot detector utilizes useful information from the parking context recognizer and does not operate when there are no parking spaces in the vicinity of the car. The coordinates of the parking slot are estimated using a two-stage deep learning model consisting of a parking context recognizer (PCR) and a parking slot detector (PSD). When the image is classified as not-parking-space, the PSD is not activated since parking is not available in that space
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
