LiDAR Sensor Data Fusion Using Bayesian Network Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing vehicle systems with multiple LiDAR sensors face challenges in aligning and fusing scan points accurately due to misalignment, which can lead to incorrect object detection and conflicting sensor readings, especially when sensors are integrated into the vehicle's fascia and cannot be physically adjusted.
Innovation Solution
A Bayesian network-based system and method for fusing outputs from multiple LiDAR sensors, which includes providing object files with position, orientation, and velocity data, constructing point clouds from scan returns, segmenting scan points into predicted clusters, matching and merging object models, creating new models, deleting dying models, and updating object files to ensure accurate alignment and fusion of data across sensors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Area of stationary object
If multiple LiDAR sensors are integrated into the vehicle's fascia, then the field-of-view coverage is improved, but sensor misalignment occurs leading to conflicting readings
Solution Approach 1:
The patent introduces a Bayesian network as an intermediary computational framework that fuses data from multiple LiDAR sensors. The system uses scan point clustering and probability calculations to reconcile discrepancies between sensors, treating the fusion algorithm as a mediator that processes conflicting readings and produces unified object detection results.
Solution Approach 2:
The system dynamically adjusts parameters such as cluster thresholds, probability cutoffs, and scan point weighting factors to optimize fusion results. By changing these parameters based on sensor characteristics and environmental conditions, the system compensates for misalignment issues while maintaining comprehensive field-of-view coverage.
2Reliability
If multiple LiDAR sensors are used to provide 360° coverage, then detection capability is improved, but data fusion complexity increases
Solution Approach 1:
The patent segments the complex fusion problem into distinct processing stages: scan point collection, cluster formation, object model generation, and Bayesian fusion. Each stage handles a specific aspect of the data, breaking down the overall complexity into manageable components that can be processed independently and then integrated.
Solution Approach 2:
The system performs preliminary clustering of scan points into clusters before the main fusion process. This preliminary organization of data reduces the computational burden of the subsequent Bayesian network operations by pre-grouping related scan points, thereby simplifying the overall fusion complexity while maintaining detection reliability.
3Productivity
If sensors track objects independently, then processing speed is improved, but conflicting readings occur due to misalignment
Solution Approach 1:
The Bayesian network implementation includes feedback mechanisms where detection results from one time step inform the processing of subsequent scan data. The system uses predicted object positions and cluster associations as feedback to guide the fusion process, allowing independent sensor processing to continue at high speed while maintaining consistency through iterative refinement of object tracks.
Data Source
AI summary
A system and method for fusing the outputs from multiple LiDAR sensors on a vehicle. The method includes providing object files for objects detected by the sensors at a previous sample time, where the object files identify the position, orientation and velocity of the detected objects. The method also includes receiving a plurality of scan returns from objects detected in the field-of-view of the sensors at a current sample time and constructing a point cloud from the scan returns. The method then segments the scan points in the point cloud into predicted clusters, where each cluster initially identifies an object detected by the sensors. The method matches the predicted clusters with predicted object models generated from objects being tracked during the previous sample time. The method creates new object models, deletes dying object models and updates the object files based on the object models for the current sample time.