Radar Point Cloud Training for Long-Range Object Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Advanced driver-assistance systems (ADAS) face challenges in accurately recognizing objects at long ranges using radar data due to lower point cloud density compared to LiDAR data, and existing neural networks have limitations in reliability and accuracy beyond a certain range.
Innovation Solution
A method involving a processor that transforms radar data into point cloud data, clusters objects based on density, determines target objects using reliability and confidence scores, and retrains a neural network using radar data to improve recognition of objects at extended ranges by generating ground truth data and tracking clusters across timepoints.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Length of stationary object
If radar data is used for object recognition at long ranges, then the measurement range is extended, but the point cloud density decreases leading to lower recognition accuracy
Solution Approach 1:
The patent introduces an intermediary processing system that uses neural networks to generate pseudo-labels from low-density radar point clouds. These pseudo-labels serve as intermediate training data to improve the neural network's ability to recognize objects at long ranges where direct radar detection is insufficient.
Solution Approach 2:
The system performs preliminary clustering and pseudo-label generation on radar data before final object recognition. By pre-processing the low-density radar point clouds to create training samples, the system prepares improved recognition models in advance that can handle long-range detection scenarios.
2Loss of time
If neural networks are trained with limited ground truth data, then training time is reduced, but the reliability and accuracy of object recognition deteriorates
Solution Approach 1:
The neural network performs self-service by generating its own training data through pseudo-labeling. The system uses the neural network itself to create labeled training samples from unlabeled radar data, eliminating the need for external manual annotation and enabling continuous self-improvement.
Solution Approach 2:
The system changes the parameter of training data quantity by generating synthetic pseudo-labeled data. This transforms the training process from being limited by scarce ground truth data to having abundant training samples, thereby improving model reliability without proportionally increasing training time.
3Ease of operation
If clustering is performed on low-density radar point clouds, then object detection is enabled, but the clustering accuracy decreases due to sparse data points
Solution Approach 1:
The patent merges multiple low-density radar point clouds over time to create denser point cloud representations. By combining data from multiple time points and using neural network-guided clustering, the system achieves accurate object detection despite the inherent sparsity of individual radar frames.
Data Source
AI summary
A method including clustering second point clouds comprised in second point cloud data of a second timepoint, obtaining output data corresponding to one or more second clusters generated as a result of the clustering by applying the second point clouds to a pretrained neural network, determining whether at least one of the one or more second clusters is a target object based on a reliability factor, tracking first clusters corresponding to the target object in first point cloud data of one or more first timepoints prior to the second timepoint, and training of the pretrained neural network based on the output data corresponding to the first clusters and the first point clouds for each first cluster determined according to a result of the tracking.