AI Urban Road Network Generation via Generative Adversarial Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional automatic generation methods for urban road networks are inefficient, particularly in new districts lacking roads, with high manual screening costs and insufficient fitting between generated and real networks, and slow model training.
Innovation Solution
An AI-based method using an interpretable generative adversarial network (infoGAN) to construct a road network rule base, incorporating specifications from urban planning, and employing UAV data and machine learning to generate multiple feasible schemes quickly, with simulation and display on a two-dimensional interaction device.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional automatic generation methods based on aerially photographed images or vehicle tracks are used, then existing roads and streets can be reproduced, but the method has limited effect for new urban districts lacking roads and requires high manual screening costs
Solution Approach 1:
The patent replaces manual screening and conventional image-based generation methods with an AI-based generative adversarial network system. The GAN automatically generates road network schemes from planning boundary data, eliminating the need for manual road drawing and screening while significantly improving design efficiency and enabling application to new urban districts without existing road data
Solution Approach 2:
The patent uses aerially photographed images and remotely sensed images as reference data to train the GAN model, allowing the system to learn and replicate realistic road network patterns. This enables automatic generation of new road networks that conform to real-world planning patterns without requiring manual intervention
2Manufacturing precision
If image learning with adversarial training is used to generate road networks in plots with strictly regulated dimensions, then a network model can be generated, but the model training speed is low and fitting between generated result and real road network is insufficient
Solution Approach 1:
The patent performs preliminary data preparation by collecting and preprocessing aerially photographed images, remotely sensed images, and vector data before training the GAN model. This preliminary action includes data cleaning, feature extraction, and creating training datasets, which accelerates the actual model training process while ensuring high fitting accuracy between generated results and real road networks
Solution Approach 2:
The patent optimizes GAN training by adjusting key parameters including learning rate, batch size, and network architecture configurations. The system also transforms input data into appropriate formats (converting vector images to bitmaps at specific resolutions) and uses multiple loss functions to improve convergence speed and fitting accuracy simultaneously
3Productivity
If multiple road network schemes are generated simultaneously using AI methods, then design efficiency is enhanced and manpower costs are reduced, but the complexity of the generation system increases
Solution Approach 1:
The patent segments the road network generation system into distinct functional modules: data acquisition module, GAN model training module, scheme generation module, and evaluation module. Each module handles specific tasks independently, which manages system complexity while enabling simultaneous generation of multiple road network schemes through parallel processing
Data Source
AI summary
The present invention discloses an artificial intelligence (AI)-based automatic generation method for an urban road network. According to the method, an anchor point distribution model is constructed by means of machine learning. Anchor points are distributed within a planning range where a boundary is a secondary trunk road. A road center line layout scheme set is generated by means of rectangular expansion. A feasible scheme set is screened out based on a rule base translated from specifications related to urban planning road, a road network scheme set is further automatically generated, and finally, a scheme is outputted to a two-dimensional interaction display device for simulated display. The present invention realizes a road network design by using a combination of machine learning and rules of the urban planning field. The present invention provides a simple and efficient automatic generation method for an urban road network. By means of the present invention, a plurality of schemes can be generated within a short time, which provide an efficient and visualized reference for the design and the practice of AI urban planning.