Point Cloud Aggregation for Non-Synchronized Sensor Fusion

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional point cloud aggregation techniques for autonomous vehicles assume static or dynamic environments, leading to smearing of dynamic or static objects and loss of contextual information when transforming sensor data to a single reference time.

Innovation Solution

A learned point cloud aggregation approach using Spatio-Temporal Graph neural networks fuses data from non-synchronized sensors into a common reference frame by transforming raw sensor data into higher-dimensional features through a learnable function, reducing artifacts caused by vehicle motion and object movement.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional point cloud aggregation transforms all sensor data to a single reference time, then temporal alignment is achieved, but per-sensor sensing artifacts and contextual information are lost

Engineering Contradiction:

Improvetemporal alignment precisionVSAvoidcontextual information loss

Core Design Contradiction:

Measurement precisionVSLoss of information

Solution Approach 1:

The patent transforms point cloud data from 3D spatial coordinates to 4D spatio-temporal features by adding temporal dimension. Each point is enhanced with temporal features (time stamp, temporal position) while preserving original spatial information, allowing simultaneous temporal alignment and context retention through multi-dimensional feature aggregation

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

Solution Approach 2:

The patent implements nested feature aggregation where original point cloud features are preserved within an outer shell of temporal features. The aggregation process nests multiple sensor readings at different timestamps into a unified spatio-temporal representation, maintaining hierarchical information structure from raw measurements to processed features

Inventive Principle:

Principle #7Nested doll (Nesting)

2Device complexity

If conventional techniques assume the world is static and adjust all points accordingly, then processing simplicity is maintained, but dynamic objects are smeared

Engineering Contradiction:

Improveprocessing complexityVSAvoidobject detection precision

Core Design Contradiction:

Device complexityVSManufacturing precision

Solution Approach 1:

The patent transitions from static world assumptions to dynamic modeling by introducing temporal features that capture motion characteristics. The system adapts to dynamic environments by learning temporal patterns from sequential sensor data, allowing differentiation between static and dynamic objects through temporal variation analysis

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The patent implements feedback mechanisms where temporal features from previous time steps inform current point cloud aggregation. The system continuously refines its understanding of object motion by comparing temporal sequences, using feedback loops to correct for vehicle motion and object dynamics iteratively

Inventive Principle:

Principle #23Feedback

3Manufacturing precision

If conventional techniques assume the world is dynamic, then dynamic objects are preserved, but static objects are smeared

Engineering Contradiction:

Improvedynamic object detection precisionVSAvoidstatic object information loss

Core Design Contradiction:

Manufacturing precisionVSLoss of information

Solution Approach 1:

The patent applies different aggregation strategies to different spatial regions based on their temporal characteristics. Static regions use aggregation methods optimized for stability while dynamic regions use methods preserving motion information. Each point cloud feature is processed with quality tailored to its specific temporal behavior pattern

Inventive Principle:

Principle #3Local quality

4Measurement precision

If learned features transform raw sensor data to higher dimensions, then aggregation accuracy is improved, but computational complexity increases

Engineering Contradiction:

Improveaggregation accuracyVSAvoidcomputational complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the point cloud aggregation process into distinct computational stages: feature extraction, temporal feature computation, and aggregation. Each stage processes data in manageable chunks, reducing peak computational complexity while maintaining overall accuracy through systematic breakdown of the transformation pipeline

Inventive Principle:

Principle #1Segmentation

Data Source

PatentUS20250355118A1Method and system for learned point cloud aggregation of non-synchronized multi-sensor fusion

Publication Date: 2025.11.20 TORC ROBOTICS INC

AI summary

A perception system is configured to: (i) initialize a grid with default values for a set of points in an environment of the perception system; (ii) based upon a first sensor data, identify a first subset of the set of points and features corresponding to a first point cloud; (iii) perform temporal alignment of the identified features corresponding to the first point cloud; (iv) update the grid using the temporally aligned features corresponding to the first point cloud; (v) based upon a second sensor data, identify a second subset of the set of points and features corresponding to a second point cloud; (vi) perform temporal alignment of the identified features corresponding to the second point cloud; and (vii) update the grid using the temporally aligned features corresponding to the second point cloud to display in a single reference frame with the temporally aligned features corresponding to the first point cloud.