3D-to-BEV Semantic Mapping for Real-Time Off-Road Traversability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing BEV mapping systems struggle to accurately project 3D sensor data into 2D maps in real-time, particularly in off-road environments, due to computational intensity and inability to distinguish between obstacles relevant and irrelevant to ground-level navigation, leading to inaccurate traversability assessments.
Innovation Solution
A Hybrid 3D to 2D Deep Convolutional Neural Network (DCNN) with a 3D convolutional encoder and 2D decoder network, utilizing skip connections and an attention-based projection layer, compresses and projects 3D data onto 2D BEV maps, filtering out irrelevant overhanging objects and upsampling to generate accurate traversability maps.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional basic projection techniques are used to convert 3D data points into 2D plane, then the projection process is simple, but the system cannot accurately distinguish relevant obstacles from irrelevant objects (such as overhanging branches) leading to reduced accuracy in traversability assessment
Solution Approach 1:
The patent segments the projection process into multiple specialized layers: a ground estimation layer that identifies ground-level surfaces, an object classification layer that distinguishes between relevant obstacles and irrelevant overhanging objects, and a BEV generation layer that creates the final map. This segmentation allows each layer to specialize in specific tasks, improving overall accuracy while managing complexity through modular design.
Solution Approach 2:
The patent introduces intermediate processing layers between the raw 3D sensor data and the final 2D BEV map. These intermediate layers include ground plane estimation modules and object relevance classification modules that act as mediators to filter and prepare data before final projection, ensuring that only relevant obstacles are included in the traversability assessment.
2Measurement precision
If computational intensive processing is applied to accurately classify objects and generate precise BEV maps, then the accuracy of traversability assessment is improved, but real-time processing capability is reduced due to increased computational load and latency
Solution Approach 1:
The patent applies partial processing by focusing computational resources only on relevant portions of the 3D data. The ground estimation layer first identifies the ground plane, and then only objects near or intersecting with the ground are classified in detail. Overhanging objects that do not intersect with the ground are processed less intensively or excluded, reducing overall computational load while maintaining accuracy for relevant obstacles.
Solution Approach 2:
The patent performs preliminary ground plane estimation and object relevance filtering before the main BEV projection and classification processes. By pre-identifying which objects are relevant to ground-level navigation, the system avoids performing full classification on all objects, thereby reducing computational latency while maintaining high accuracy for relevant obstacles.
3Loss of information
If all 3D data points are projected into 2D BEV map without filtering, then no information is lost, but the map contains irrelevant objects (such as overhanging branches) that reduce navigation safety and increase processing complexity
Solution Approach 1:
The patent extracts and removes irrelevant objects from the BEV map through a filtering process. After projecting 3D data into 2D space, the system identifies objects that do not intersect with the ground plane (such as overhanging branches) and excludes them from the final traversability map. This extraction process maintains completeness of relevant terrain information while eliminating harmful false obstacles that would compromise navigation safety.
4Manufacturing precision
If high resolution 3D sensor data is processed to generate detailed BEV maps, then the detail and accuracy of terrain representation is improved, but the computational resources and processing time required increase significantly
Solution Approach 1:
The patent applies local quality processing by varying the level of detail processed for different regions of the 3D data. Areas near the robot and containing relevant obstacles are processed at high resolution, while distant areas or regions containing only irrelevant objects are processed at lower resolution or excluded entirely. This selective processing maintains high terrain representation detail where needed while reducing overall computational energy consumption.
Data Source
AI summary
A mapping system for an autonomous mobile robot includes a 3D convolutional encoder network that generates 3D feature maps from 3D point cloud data. The network sequentially compresses the feature dimension of the 3D input data to reduce the computational complexity and enable feature extraction to be performed in substantially real-time. Skip connections connect the outputs of the encoder layers of the convolutional encoder network to counterpart decoder layers of a 2D convolutional decoder network. An attention-based 3D to 2D projection layer receives the 3D feature maps generated by the encoder layers via the skip connections and projects the 3D feature maps onto 2D BEV feature maps which are provided to the counterpart decoder layers as input. The projection layer automatically estimates ground level of 3D feature maps and filters out overhanging objects that are irrelevant to ground-level navigation.