Fast Parking Slot Detection in the Bird's Eye View

Vision-based parking slot detection plays an important role for autonomous vehicles to achieve automatic parking. Complex visual environments severely affect the accuracy of parking slot detection and occupancy classification, such as light, weather, shadows, and ground textures and so on. To solve this problem, we propose a deep learning-based fast parking slot detection method in the bird's eye view image, namely FPS-Net. Firstly, given a bird's eye view, a parking slot detection method based on MobileNetv3 is proposed to predict the location, shape and orientation of the marking points. Secondly, the four corner points of the parking slot are inferred by post-processing. Finally, a parking slot is determined as vacant or not based on the distribution of extracted features using HOG feature extraction. From the experimental results it can be seen that the FPS-Net can identify various types of parking slots with an average precision of 98.34% in the PS2.0 dataset and achieve 87.39% accuracy for occupation classification.