LiDAR Object Identification Using Multi-View Neural Network Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing LiDAR systems face challenges in efficiently identifying and classifying objects in a vehicle's surroundings, requiring significant computational resources and often failing when a subset of neural networks is not functioning as expected.
Innovation Solution
The use of parallel training and deployment of separate view neural networks (BeV and RV) for LiDAR data processing, combined with a recurrent layer for spatial smoothing and data augmentation, enhances object classification accuracy by fusing multiple views of a LiDAR scan.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple neural networks are used for LiDAR data processing to improve object classification accuracy, then classification accuracy is improved, but computational resources required increase
Solution Approach 1:
The patent divides the LiDAR data processing into multiple separate view neural networks (BeV and RV networks) that process different views of the point cloud data independently. Each network handles a specific view type, and their results are fused to achieve accurate classification. This segmentation allows parallel processing and reduces the computational burden on a single network while maintaining high classification accuracy.
Solution Approach 2:
The patent combines the outputs from multiple separate view neural networks through a fusion mechanism. The BeV network and RV network results are merged to produce the final object classification, leveraging the complementary information from different views. This merging approach improves accuracy by utilizing multiple perspectives while distributing computational work across separate networks.
2Reliability
If a single neural network is used for LiDAR processing, then device complexity is reduced, but reliability decreases when the network fails
Solution Approach 1:
The system is segmented into multiple independent view neural networks (BeV and RV networks) that can operate autonomously. If one network fails or produces unreliable results, the other network can still provide classification output, ensuring system reliability. The independence of each network module allows for fault tolerance without requiring complete system failure.
Solution Approach 2:
The patent employs data augmentation techniques that transform the input LiDAR point cloud data through various parameter changes (rotations, scaling, noise addition) to create diverse training samples. This enables the neural networks to learn more robust features and maintain reliable performance under varying conditions, improving overall system reliability through enhanced generalization capability.
3Measurement precision
If parallel training of multiple view neural networks is implemented, then object classification accuracy is improved through multi-view fusion, but training time and computational load increase
Solution Approach 1:
The training process is segmented into parallel training of separate view neural networks. Each network (BeV and RV) is trained independently on its respective view type, allowing simultaneous training execution. This parallelization reduces the overall training time compared to sequential training while maintaining the ability to fuse results for accurate classification.
Solution Approach 2:
The patent applies data augmentation to generate additional training samples by applying transformations to the original LiDAR data. This creates an expanded training dataset that improves network robustness and accuracy without requiring proportionally more training time, as the augmented samples are generated computationally rather than requiring additional real-world data collection.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach reduces computational requirements and improves object classification accuracy by leveraging multiple neural networks to handle uncertainties, resulting in more precise object identification for autonomous vehicles.
Implementation Method 1
LiDAR is a technology that uses light to obtain data about physical objects in line of sight of a light emitter
Data Source
AI summary
Among other things, techniques are described for controlling, using a control circuit, motion of a vehicle based objects identified using LiDAR. For example, respective classes of points of a point cloud are determined, and based on the determined respective classes of the points of the point cloud, objects in the vicinity of the vehicle are identified.