RGB, Infrared, and Depth Fusion for Driver Action Recognition

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing driver action recognition (DAR) methods primarily rely on single-modal sensors, which are sensitive to lighting conditions and environmental variations, leading to suboptimal performance due to complex cabin environments and limited driver body movements, and fail to effectively leverage inter-relations between multi-modal sensor inputs.

Innovation Solution

A multi-modal fusion transformer model, named MultiFuser, employs multi-stream encoders to extract intra-modal features and cross-modality fusion decoders to integrate inter-modal features, using focal fusion blocks and multi-modal consistent attention to enhance feature representation and computational efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If single-modal sensors are used for driver action recognition, then the system is simpler and less computationally intensive, but the performance deteriorates due to sensitivity to lighting conditions and environmental variations

Engineering Contradiction:
Improveaction recognition performanceVSAvoidsensor system complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent merges multiple modalities (RGB, infrared, depth) into a unified multi-modal fusion transformer architecture. The encoder processes all modalities through shared transformer blocks that perform cross-attention mechanisms, allowing the system to leverage complementary information from different sensors to improve recognition reliability under varying lighting and environmental conditions.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The transformer architecture serves multiple functions simultaneously: it encodes temporal dependencies, performs cross-modality attention, and generates action predictions. This multi-functional design allows a single system to handle diverse sensor inputs and complex action recognition tasks without requiring separate specialized processing pipelines for each modality.

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Reliability

If multi-modal fusion is implemented to improve recognition accuracy, then the feature representation becomes more comprehensive, but the computational cost increases

Engineering Contradiction:
Improveaction recognition accuracyVSAvoidcomputational energy consumption
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent segments the computational graph into distinct encoder and decoder portions, where the encoder processes all modalities in parallel through shared transformer blocks, and the decoder generates predictions from the encoded representations. This segmentation allows efficient memory access patterns and enables gradient accumulation across modalities without requiring simultaneous processing of all data at once.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The model dynamically adjusts the number of transformer blocks and attention heads based on the input modalities and task requirements. The architecture can be configured with varying numbers of encoder/decoder layers and attention mechanisms, allowing optimization of the computational-to-performance ratio for different application scenarios.

Inventive Principle:
Principle #35Parameter changes

3Productivity

If deep learning architectures focus on single-modality feature extraction, then the model is simpler and trains faster, but the performance is insufficient for complex action recognition scenarios

Engineering Contradiction:
Improvetraining speedVSAvoidaction recognition capability
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent merges single-modality encoder blocks into a multi-modal architecture where shared transformer blocks process multiple modalities simultaneously. This allows the model to learn cross-modality relationships and comprehensive action representations while maintaining the training efficiency of standardized transformer components that can be pre-trained and fine-tuned.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The encoder performs preliminary feature extraction and representation learning for all modalities before the prediction head processes the combined representations. This preliminary encoding stage can leverage pre-trained weights and efficient transformer computations to quickly establish robust feature representations that improve both training speed and recognition capability.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4641522A1Method and system for action recognition with multi-modal fusion
Publication Date: 2025.10.29 CONTINENTAL AUTOMOTIVE TECHNOLOGIES GMBH
  • EP4641522A1 patent drawingFigure 1
  • EP4641522A1 patent drawingFigure 2
  • EP4641522A1 patent drawingFigure 3~4

AI summary

The present disclosure relates to a computer-implemented method for action recognition, the method comprising: a) capturing a plurality of sequence of frames wherein each sequence is recorded using a predetermined modality; b) extracting intra-modality features and deriving a class token using N layers of multi-stream encoder; c) passing intra-modality features through K layers of cross-modality fusion decoder to extract inter-modality features and a query token; d) concatenating the class token with the query token to generate a global representation; and e) feeding the global representation to an action classification module trained to determine an action performed based on the global representation, and determining the action that was most probably performed. The present disclosure further relates to a control method for a vehicle, a system, a vehicle, a computer program, and a computer readable medium or data carrier signal.