Grid Map Generation Using Dynamic Object Filtering for Low-Cost Navigation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing robot navigation technologies are costly due to the expense of lidar and wheel odometers, and it is challenging to generate a grid map for navigation using inexpensive alternatives.
Innovation Solution
A grid map generation method that uses a preset learning model, such as a truncated signed distance function (TSDF) model, to process original image data and pose data from a photographing device, performing dynamic object filtering and fusion to create an obstacle-free grid map, reducing generation costs by utilizing low-cost imaging and pose data acquisition devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If lidar and wheel odometer are used for global map reconstruction and grid map generation, then the navigation accuracy and reliability are improved, but the system cost increases significantly
Solution Approach 1:
The patent replaces expensive lidar and wheel odometer with inexpensive photographing devices (cameras) and pose sensors. The camera captures images that are processed through learning models to generate grid maps, while the pose sensor provides low-cost pose data. This substitution dramatically reduces system cost while maintaining navigation functionality through software-based processing of visual data.
Solution Approach 2:
The patent substitutes mechanical sensing systems (lidar's light detection, wheel odometer's mechanical rotation sensing) with a vision-based system. The photographing device captures optical information that is then processed through TSDF learning models and dynamic object filtering algorithms to reconstruct 3D space and generate navigation maps, replacing physical mechanical sensing with optical sensing and computational processing.
2Device complexity
If inexpensive photographing devices and pose sensors are used instead of lidar and wheel odometer, then the system cost is reduced, but the ability to generate accurate grid maps for navigation deteriorates
Solution Approach 1:
The patent performs preliminary actions by pre-training learning models with TSDF (Truncated Signed Distance Function) and performing dynamic object filtering before actual grid map generation. The system pre-processes image data through learned models that have been trained to recognize and filter dynamic objects, ensuring that the subsequent grid map generation starts with cleaned, reliable data that compensates for the lower quality of inexpensive sensors.
Solution Approach 2:
The patent introduces learning models as intermediaries between the inexpensive photographing devices and the final grid map output. The TSDF-based learning model acts as a mediator that processes raw image data from the camera and pose data from the sensor, performing dynamic object filtering and 3D reconstruction to transform low-quality sensor inputs into high-quality navigation-ready grid maps.
3Manufacturing precision
If dynamic object filtering is performed using learning models, then the impact of dynamic objects on grid map accuracy is reduced, but the processing time and computational complexity increase
Solution Approach 1:
The patent performs dynamic object filtering as a preliminary action before final grid map generation. By using pre-trained learning models to filter out dynamic objects from image data in advance, the system ensures that the subsequent grid map reconstruction processes only static environmental features, significantly improving grid map precision while managing processing time through efficient pre-processing.
Data Source
AI summary
The embodiments of the present disclosure relate to the field of communication technology, and disclose a grid map generation method and device, and a computer-readable storage medium. The grid map generation method includes: acquiring original image data of an original image photographed by a photographing device and pose data of the photographing device when photographing the original image; inputting the original image data and the pose data into a preset learning model to obtain first image data, wherein the preset learning model performs a first dynamic object filtering on the original image data according to the pose data to obtain the first image data; fusing the first image data and the pose data into the preset learning model to obtain a fusion learning model; and obtaining a grid map of the original image data according to the fusion learning model and the pose data.