LiDAR Pose Estimation via Voxelized Point Clouds
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing pose estimation techniques using camera sensors face challenges in low-light environments and struggle to accurately determine the pose of individuals due to the lack of depth information, making them unsuitable for autonomous driving applications.
Innovation Solution
The use of LiDAR sensors to generate point clouds, which provide 3D maps with depth information, combined with deep neural networks that process voxelized frames using both 3D and 2D convolutional layers to detect persons and estimate their poses, enabling accurate and efficient pose estimation even in low-light conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Illumination intensity
If camera sensors are used for pose estimation, then the system can operate in well-lit environments, but it fails in low-light environments and lacks depth information
Solution Approach 1:
The patent replaces camera-based optical sensing with LiDAR-based active sensing. LiDAR uses laser pulses to actively illuminate and measure the environment, substituting passive camera sensing that fails in low-light conditions. This enables depth information acquisition and pose estimation to function independently of ambient illumination levels.
Solution Approach 2:
The patent changes the fundamental sensing parameter from 2D image intensity (camera) to 3D spatial point cloud coordinates (LiDAR). By measuring distance through time-of-flight of laser pulses rather than detecting reflected light intensity, the system gains depth information and operates effectively in low-light environments where camera sensors fail.
2Measurement precision
If 3D convolutional layers are used for processing voxelized frames, then measurement precision improves, but processing time increases
Solution Approach 1:
The patent segments the deep neural network into two distinct processing stages: first applying 3D convolutional layers to extract spatial features and detect persons with high accuracy, then applying 2D convolutional layers for refined pose estimation. This segmentation allows each layer type to specialize in specific tasks, balancing accuracy and processing efficiency.
Solution Approach 2:
The patent transitions from processing raw 3D point cloud data directly to first converting data into a voxelized 3D grid representation, then processing through both 3D and 2D convolutional layers. This dimensional transformation enables efficient feature extraction while maintaining accuracy, as the structured voxel format facilitates both spatial and planar pattern recognition.
Data Source
AI summary
An apparatus and method for estimating the pose of one or more persons is described. An example technique may include receiving a point cloud from a LiDAR sensor, the point cloud including a plurality of points representing positions of objects relative to the LiDAR sensor, processing the point cloud to produce a voxelized frame including a plurality of voxels, processing the voxelized frame using a deep neural network to determine one or more persons relative to the LiDAR sensor and a pose for each of the one or more persons, and outputting a location of the determined one or more persons and the pose for each of the determined one or more persons.