LiDAR Ground Plane Estimation Using 2D Pillar Semantic Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing autonomous vehicle perception pipelines face challenges in accurately estimating the ground plane using LiDAR data, which is crucial for efficient route planning and autonomous driving.
Innovation Solution
A LiDAR semantic network (LSN) is employed, which encodes LiDAR point clouds into 2D pillars, processes them through a convolutional backbone to generate feature maps, and uses a ground plane estimation head to accurately estimate the ground plane, enabling drivable area determination and path planning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning techniques are used to classify LiDAR points as foreground or ground, then the perception pipeline is simpler, but the ground plane estimation accuracy is insufficient
Solution Approach 1:
The patent transforms the 3D point cloud data into a 2D bird's eye view representation with height encoding. This dimensional transformation allows the use of 2D convolutional neural networks for 3D ground plane estimation, simplifying the processing while maintaining accuracy. The point cloud is projected onto a 2D plane where each pixel represents a vertical column of points, enabling efficient processing with standard 2D CNN architectures.
Solution Approach 2:
The patent changes the parameter representation by encoding height information into the 2D image space. Instead of processing raw 3D coordinates directly, the system transforms spatial coordinates into a 2D representation where height becomes an additional dimension in the image space, allowing traditional 2D image processing techniques to be applied effectively to 3D ground plane estimation.
2Measurement precision
If deep learning networks with multiple strides are used to estimate the ground plane, then the estimation accuracy is improved, but the computational complexity increases
Solution Approach 1:
The patent divides the feature extraction process into multiple stages with different strides. The convolutional neural network processes the 2D bird's eye view image at multiple resolution levels, with each stride capturing features at different scales. This segmentation allows the system to focus computational resources efficiently at each level, achieving high accuracy without requiring excessive computational power at any single stage.
Solution Approach 2:
The patent employs a dynamic multi-stride architecture where the network adaptively processes features at different resolution levels. The system dynamically adjusts the processing depth and stride at different regions of the image, allocating more computational resources to areas requiring higher precision while reducing processing in less critical regions, thereby optimizing overall computational efficiency.
Data Source
AI summary
Embodiments are disclosed for ground plane estimation (GPE) using a LiDAR semantic network. In an embodiment, a method comprises: obtaining a point cloud from a depth sensor of a vehicle operating in an environment; encoding the point cloud; estimating, using a deep learning network with the encoded point cloud as input, a ground plane in the environment; planning a path through the environment based on a drivable area of the estimated ground plane; and operating the vehicle, the vehicle along the path. The deep learning network includes a two-dimensional (2D) convolutional backbone, a detection head for detecting objects and a GPE head for estimating the ground plane. In an embodiment, point pillars are used to encode the point cloud.