3D Point Cloud Temporal Aggregation for Online Object Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current 3D object detection methods for indoor scenes struggle with online detection, relying on spatial relationships and consuming excessive computational resources due to redundant computations, and fail to accurately identify both large and small objects in dynamic environments.

Innovation Solution

A neural network-based method that trains on sequences of 3D point cloud frames, leveraging spatial and temporal relationships to output localized representations of objects by aggregating feature vectors from previous frames, reducing computational overhead while maintaining accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If temporal aggregation is applied to improve online detection accuracy, then detection quality improves, but computational complexity increases

Engineering Contradiction:
Improvedetection accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The neural network processes point cloud data by segmenting it into multiple resolution levels (coarse and fine resolutions). The aggregation operation combines features from different resolutions and temporal frames in a hierarchical manner, reducing computational complexity while maintaining detection accuracy. This segmentation approach allows the system to process temporal sequences efficiently without computing all possible combinations.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a temporal dimension to the detection process by aggregating features across multiple frames. Instead of processing each frame independently, the system adds the time dimension to accumulate information from previous frames, improving detection accuracy for dynamic scenes while using efficient aggregation operations to manage computational load.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If temporal aggregation is applied to improve online detection accuracy, then detection quality improves, but resource consumption increases

Engineering Contradiction:
Improvedetection accuracyVSAvoidcomputational resource consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

By segmenting the feature aggregation into resolution-level stages (coarse then fine), the system reduces redundant computations. The coarse-resolution aggregation provides a rough temporal context that guides the finer-resolution processing, thereby reducing overall computational resource consumption while maintaining detection accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system performs partial aggregation by selectively combining features from previous frames based on detection needs. Rather than processing all possible feature combinations, the neural network applies aggregation operations only where necessary, optimizing resource consumption while achieving sufficient detection accuracy for real-time applications.

Inventive Principle:
Principle #16Partial or excessive action

3Measurement precision

If spatial relationships are emphasized in detection, then object localization improves, but temporal dynamics are neglected

Engineering Contradiction:
Improveobject localization accuracyVSAvoidtemporal adaptability
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent adds the temporal dimension to the spatial detection framework by incorporating feature aggregation across multiple frames. The neural network maintains spatial relationships for accurate localization while simultaneously capturing temporal dynamics through sequential frame aggregation, enabling the system to adapt to moving objects and changing scenes.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Solution Approach 2:

The system merges spatial and temporal information processing by combining feature extraction from multiple frames with spatial relationship analysis. The aggregation operation integrates temporal context with spatial localization, allowing the detector to maintain accurate object positioning while adapting to temporal changes in the scene.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP4654152A1Temporal aggregation for online 3D object detection
Publication Date: 2025.11.26 DASSAULT SYSTEMES SA
  • EP4654152A1 patent drawingFigure 1~2
  • EP4654152A1 patent drawingFigure 3~4
  • EP4654152A1 patent drawingFigure 5

AI summary

The disclosure notably relates to a computer-implemented method for machine learning a function. The function is configured to take as input a 3D point cloud frame of a real scene and to output localized representations. Each output representation is respective to a respective object of the real scene. The method comprises obtaining a dataset of sequences of 3D point cloud frames. Each frame is associated with a time in the sequence. Each frame comprises localized representations each of a respective object. The method further comprises training the function based on the obtained dataset. The training comprises, for each sequence of the dataset and each given frame of the sequence, training the function to output localized representations of objects in the given frame based on the given frame and on at least the frame with the previous time in the sequence.