Channel-Specific Convolution for 3D LiDAR Object Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing convolutional neural network (CNN) methods for 3D LiDAR object recognition face data loss and increased computation time due to the need for converting point cloud data into voxels or images, which also require additional algorithms for accurate 3D posture estimation.
Innovation Solution
A convolutional neural network structure based on channel-specific convolution that directly processes raw LiDAR data, using an inside-channel network for feature extraction within channels and an outside-channel network for feature extraction between channels, eliminating the need for data conversion and reducing computational dimensions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If point cloud data is converted into voxel for 3D convolution, then object recognition accuracy is improved, but computation time increases significantly
Solution Approach 1:
The patent segments the LiDAR data processing by dividing channels into inside-channel and outside-channel components. Inside-channel convolution processes each channel independently with reduced dimensions, while outside-channel convolution handles inter-channel relationships. This segmentation avoids full 3D convolution on voxel data while maintaining recognition accuracy.
Solution Approach 2:
The patent transforms the processing approach by operating in a channel-based dimensional space rather than converting to full 3D voxel space. The channel-specific convolution extracts features in the original LiDAR channel dimension, avoiding the computational burden of 3D convolution while preserving spatial information through the channel structure.
2Productivity
If point cloud data is converted into image for 2D processing, then computation speed is improved, but 3D posture estimation accuracy deteriorates
Solution Approach 1:
The patent utilizes the channel dimension of LiDAR data as an intermediate dimensional space that preserves 3D spatial information without requiring full 3D voxel conversion. The channel-specific convolution operates in this dimensional space to extract features that maintain 3D posture information while avoiding the computational complexity of 3D convolution.
Solution Approach 2:
The patent applies different convolution operations to different parts of the data: inside-channel convolution processes local features within each channel, while outside-channel convolution processes global features across channels. This local quality approach allows efficient processing while preserving the 3D spatial relationships needed for accurate posture estimation.
3Ease of manufacture
If point cloud data is converted into voxel or image, then data processing becomes feasible, but data loss occurs
Solution Approach 1:
The patent extracts essential features directly from the raw LiDAR point cloud data through channel-specific convolution, avoiding the information loss that occurs during voxel or image conversion. By operating on the original data structure and extracting features in the channel dimension, the method preserves the完整性 of the original data while making it computationally tractable.
Data Source
AI summary
Provided is a convolutional neural network structure based on channel-specific convolution for high-speed object recognition of a 3D LiDAR, including: an inside-channel convolutional network for extracting features in channels of a LiDAR data; an outside-channel convolutional network for extracting features between the channels by using outputs for the channels of the inside-channel convolutional network and generating a feature map representing the extracted features between the channels; and a detection network searching location and class of an object by using the feature map generated from the outside-channel convolutional network. According to the invention, since a raw data coming from a LiDAR is directly used in a high-speed object recognition procedure of a 3D LiDAR, there is an effect that data loss does not occur.