Camera-Based Parking Space Detection in Unmarked Lots
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing parking assist systems face challenges in accurately identifying and navigating into parking spaces, especially when parking lines are unclear or absent, and when encountering unfamiliar parking lots not included in the training data of spatial recognition models.
Innovation Solution
The method involves using cameras on a moving object to capture images of its surroundings, performing object detection to identify candidate parking areas, and applying scene segmentation to determine if these areas are occupied. Based on these analyses, the system determines whether the moving object can fit into the candidate area using a template corresponding to its size.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional parking assist systems use sensors and spatial recognition models to identify parking spaces, then parking automation is achieved, but the system fails when parking lines are unclear or absent and when encountering unfamiliar parking lots
Solution Approach 1:
The patent replaces traditional mechanical/sensor-based parking space detection (relying on clear parking lines and pre-trained spatial recognition models) with a vision-based approach using large language models and image generation technology. The system captures images of the parking environment, generates synthetic images with annotated parking spaces, and uses these to train and adapt to unfamiliar parking lots without requiring clear physical markings or pre-existing spatial models.
2Measurement precision
If the system relies on clear parking lines and trained spatial recognition models, then measurement precision is improved, but the system cannot handle unfamiliar parking lots not included in training data
Solution Approach 1:
The system performs preliminary actions by capturing images of the parking environment and generating synthetic training data before actual parking space identification. It creates annotated images with parking spaces marked, trains the spatial recognition model on this generated data, and then uses the trained model to accurately identify parking spaces in the actual environment. This preliminary data generation and model training enables the system to achieve high measurement precision in unfamiliar parking lots.
3Adaptability or versatility
If the system uses image-based object detection and scene segmentation, then adaptability to unfamiliar environments is improved, but the complexity of image processing increases
Solution Approach 1:
The patent introduces an intermediary image generation process that bridges the gap between raw camera images and parking space identification. The system generates synthetic images with annotated parking spaces as an intermediate representation, which then serves as training data for the spatial recognition model. This intermediary step simplifies the overall process by creating a standardized, annotated dataset that the model can efficiently process, reducing the complexity of direct image-to-parking-space mapping.
Data Source
AI summary
A method and device with parking space navigation are provided. An operating method of a moving object includes: obtaining, from cameras of the moving object, images of surroundings of the moving object that are captured by the cameras; determining a candidate area for parking the moving object by performing object detection on the images; determining whether the candidate area is occupied by performing scene segmentation on the images; and based on determining that the candidate area is not occupied, determining whether the moving object is able to be parked into the candidate area based on a template area corresponding to a size of the moving object.