Hybrid LiDAR Segmentation for Semantic Outlier Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for LiDAR-based autonomous driving struggle to segment real-world objects effectively, as closed world approaches fail to handle unseen data, while open world approaches lack semantic meaning for identified clusters, posing safety concerns for autonomous vehicle interactions.
Innovation Solution
A hybrid approach that segments point clouds using a neural network trained with a first loss function for known segments and a second loss function for outlier segments, combined with an extended isolation forest to generate anomaly scores, providing semantic classification and outlier detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a closed world approach is used for LiDAR segmentation, then the model can be trained on a dataset with known objects, but the model fails to deal with unseen data and objects
Solution Approach 1:
The patent segments the segmentation task into two distinct parts: a closed-world neural network for classifying known objects and an open-world isolation forest for detecting unknown objects. This segmentation allows each component to specialize - the neural network achieves high accuracy for known classes while the isolation forest provides adaptability for unseen classes, resolving the contradiction between precision and versatility.
Solution Approach 2:
The patent introduces an intermediary mechanism - the isolation forest - that acts as a bridge between the closed-world neural network and the open-world detection requirement. The isolation forest receives features from the neural network and identifies outliers representing unseen objects, thereby mediating between the limited scope of trained classes and the need for generalization to unknown objects.
2Adaptability or versatility
If an open world approach is used to identify all possible clusters, then the system can detect unseen objects, but the identified clusters lack semantic meaning
Solution Approach 1:
The patent segments the processing pipeline into feature extraction (neural network) and outlier detection (isolation forest). The neural network preserves semantic information by classifying points into meaningful categories, while the isolation forest operates on these semantically-enriched features to detect unknown objects. This segmentation ensures that semantic meaning is not lost in the open-world detection process.
Solution Approach 2:
The patent creates a composite approach combining two different methodologies - the semantic-rich neural network classification and the anomaly-detection isolation forest. By compositing these two approaches, the system maintains the semantic strengths of the neural network while adding the adaptability of the isolation forest for detecting unseen objects, preventing loss of semantic information.
3Measurement precision
If only known segments are classified by the neural network, then the classification is accurate for training data, but outlier segments cannot be properly identified
Solution Approach 1:
The isolation forest serves as an intermediary safety layer between the neural network classification and autonomous vehicle decision-making. It detects outlier segments that the neural network may misclassify, providing an additional reliability check. This intermediary mechanism ensures that unseen objects do not lead to unsafe interactions, addressing the reliability concern while preserving neural network accuracy for known objects.
Solution Approach 2:
The patent applies beforehand cushioning by using the isolation forest to pre-identify potential outliers before they reach the autonomous vehicle's decision-making system. This prior detection creates a safety buffer, cushioning against the risks of misclassifying unseen objects and allowing the system to prepare appropriate safety responses in advance.
4Ease of manufacture
If a single loss function is used for training, then the training process is simple, but the model cannot optimize both known and outlier segment classification
Solution Approach 1:
The patent segments the training process into two distinct training phases: one for the neural network using classification loss for known objects, and another for the isolation forest using anomaly detection loss for outliers. This segmentation of the training process allows each component to be optimized for its specific function while maintaining overall system performance, resolving the contradiction between training simplicity and classification precision.
Data Source
AI summary
Devices, methods, and systems may obtain at least one point cloud, segment points in the at least one point cloud into a plurality of segments, train a neural network using known segments and a first loss function to generate a first trained neural network, train the first trained neural network using outlier segments and a second loss function to generate a second trained neural network, and train an extended isolation forest by applying an extended isolation algorithm to features of the known segments and features of the outlier segments to generate an anomaly score for each segment.