Radar Map Generation From Aerial Images for Autonomous Navigation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for generating maps for autonomous vehicles are expensive, time-consuming, and require constant updating due to environmental changes, especially when relying on vehicles dedicated to mapping with sensors.
Innovation Solution
The use of trained deep learning models to process aerial image data and generate predicted radar maps, which include residential, highway, suburban, and rural models, to control vehicles effectively, optimizing hyper-parameters like number of layers, filter size, and class weights for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If vehicles dedicated to mapping with sensors are used to generate maps, then map data can be collected, but the process becomes expensive and time consuming
Solution Approach 1:
The patent uses aerial images as a copy or alternative source of spatial information instead of relying on dedicated mapping vehicles. The deep learning model processes these aerial images to generate radar-like maps, effectively copying the mapping function from physical sensor vehicles to computational processing of available aerial data, thereby reducing time and cost while maintaining map quality
Solution Approach 2:
The patent replaces the mechanical system of dedicated mapping vehicles physically traversing areas with a computational system that processes aerial images using deep learning. This substitution eliminates the need for physical mapping vehicles to collect data, dramatically reducing time consumption and operational costs while preserving the ability to generate accurate environmental maps
2Reliability
If maps are constantly updated to reflect environmental changes, then navigation accuracy is maintained, but the cost and time requirements increase
Solution Approach 1:
The system enables self-updating of maps by processing newly available aerial images through the trained deep learning model. When environmental changes occur, new aerial imagery can be automatically processed to regenerate updated radar maps without requiring dedicated mapping vehicles to revisit the area, allowing the system to self-maintain navigation accuracy efficiently
Solution Approach 2:
The deep learning model is pre-trained on diverse aerial images and radar data, enabling it to quickly adapt and process new aerial imagery when environmental changes occur. This preliminary preparation allows rapid map updates without extensive processing time, maintaining navigation reliability while improving updating efficiency
3Adaptability or versatility
If multiple deep learning models are used for different environments, then navigation accuracy across diverse locations is improved, but system complexity increases
Solution Approach 1:
The patent segments the mapping problem by creating specialized deep learning models for different environmental types (residential, highway, suburban, urban, rural). Each model is optimized for its specific environment, improving navigation accuracy in diverse locations. The system manages this complexity by organizing models into distinct categories and selecting appropriate models based on the detected environment type
4Measurement precision
If deep learning models are trained with optimized hyper-parameters, then map generation accuracy is improved, but training time and computational resources increase
Solution Approach 1:
The patent performs preliminary optimization of hyper-parameters during the model training phase, establishing optimal settings for number of layers, filter size, filter depth, class weights, and number of epochs. This preliminary action ensures that once the models are deployed for map generation, they achieve high accuracy efficiently without requiring repeated training adjustments, balancing training investment with operational performance
Data Source
AI summary
Systems and methods are provided for generating a map for use in controlling a vehicle. In one embodiment, a method includes: receiving, by a processor, aerial image data depicting an environment; processing, by the processor, the aerial image data with a plurality of trained deep learning models to produce a predicted radar map; and controlling the vehicle based on the predicted radar map.