Obstacle Detection Using Self-Supervised Masks and LiDAR Clusters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current autonomous driving perception systems struggle to detect general obstacles not included in predefined categories due to variability and edge cases, and supervised learning methods fail to address open-world detection effectively.
Innovation Solution
A method combining foundational models like Grounding DINO and Segment Anything for self-supervised obstacle segmentation, combined with computational geometry-based outlier detection using sensors like LiDAR, to identify and classify obstacles without explicit supervision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning with fixed categories is used for obstacle detection, then detection accuracy for known obstacles is improved, but the system cannot detect general obstacles not included in the predefined category set
Solution Approach 1:
The system employs self-supervised learning where the model learns from unlabeled data automatically without requiring manual annotation. The segmentation model identifies obstacles and their categories autonomously, enabling the system to adapt to new obstacle types without external supervision or retraining with labeled data
Solution Approach 2:
The system transitions from static fixed categories to dynamic open-world detection. The segmentation model can adaptively identify and classify arbitrary obstacle types in real-time based on visual features, allowing the detection capability to evolve dynamically without being constrained by predefined categories
2Adaptability or versatility
If extensive training with labeled data is performed to improve detection coverage, then more obstacle categories can be detected, but the system complexity and training time increase significantly
Solution Approach 1:
The system uses self-supervised learning to automatically learn from unlabeled data without requiring manual annotation efforts. The model performs its own training by identifying patterns in raw visual data, eliminating the need for extensive labeled datasets and complex training pipelines while maintaining high detection coverage
Solution Approach 2:
The system leverages pre-trained foundational models that have been trained on large-scale datasets. By copying and adapting these pre-trained models to the obstacle detection task, the system benefits from existing knowledge without requiring retraining from scratch, significantly reducing training complexity and time
3Speed
If traditional segmentation networks are used for obstacle detection, then processing speed is maintained, but the system lacks the capability to detect unknown categories lying on the road surface
Solution Approach 1:
The system replaces traditional static segmentation networks with dynamic vision transformers that can adapt to new obstacle categories. The transformer architecture processes visual inputs through self-attention mechanisms, enabling flexible identification of unknown obstacles while maintaining real-time processing speeds through optimized computational operations
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present disclosure relates to a method and system for obstacle detection in autonomous driving. The methods includes: receiving, by a processing module, first road information from a first sensing module; determining, by the processing module, a set of obstacle masks based on the first road information by applying a self-supervised method; receiving, by the processing module, second road information from a second sensing module; determining, by the processing module, a set of clusters based on the second road information by applying an clustering method; and determining, by the processing module, an obstacle based on the set of obstacle masks and the set of clusters.