Driver Behavior Prediction Using Physiological and Road Scene Fusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional Transformer models for human-factor intelligent driving behavior prediction exhibit high adaptability yet low accuracy, poor interpretability, significant time consumption due to separate feature extraction in sub-task modules, and reduced real-time performance when used in intelligent cabin environments.
Innovation Solution
A method involving physiological signal analysis, multi-period decomposition, two-dimensional spatial expansion, multi-modal data fusion, and three-dimensional backbone network processing to enhance driving behavior prediction, including feature analysis and inference layers for real-time performance improvement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the Transformer model is used for driving behavior prediction, then adaptability is improved, but prediction accuracy deteriorates
Solution Approach 1:
The patent segments the driving behavior prediction task into multiple sub-tasks including lane change prediction, acceleration prediction, and steering prediction. Each sub-task is handled by a dedicated prediction module within the Transformer architecture, allowing the model to specialize in specific aspects of driving behavior while maintaining overall adaptability through the unified framework.
2Adaptability or versatility
If the end-to-end autonomous driving model based on Transformer is used, then model generalization is improved, but interpretability deteriorates
Solution Approach 1:
The model is divided into distinct functional modules including encoding module, prediction module, and decoding module. Each module has a clear function and can be independently analyzed, which improves interpretability while maintaining the generalization capabilities of the end-to-end Transformer architecture through the modular design.
3Reliability
If separate feature extraction sub-task modules are employed for different sensors, then processing comprehensiveness is improved, but time consumption deteriorates
Solution Approach 1:
The patent merges the feature extraction processes for different sensor types (visual data, radar data, etc.) into a unified Transformer model architecture. Instead of employing separate deep learning models for each sensor type, the model processes multi-modal data together through shared encoding and attention mechanisms, reducing redundant computations and time consumption while maintaining comprehensive processing capabilities.
4Reliability
If complete segment of data is used for each prediction task, then prediction completeness is improved, but real-time performance deteriorates
Solution Approach 1:
The patent implements a sliding window mechanism that performs preliminary processing on incoming data streams. The model maintains a buffer of recent historical data and continuously updates predictions based on new inputs without waiting for complete data segments, enabling real-time performance while ensuring prediction completeness through the accumulation of sufficient historical context in the sliding window.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Improves the accuracy and real-time performance of driving behavior prediction by reducing time consumption in sub-task modules and enhancing feature extraction, while maintaining model generalization and interpretability.
Implementation Method 1
performing fast Fourier transform on the physiological signal to generate an amplitude-frequency characteristic
Data Source
Figure 1
Figure 2~3
Figure 4~5
AI summary
The embodiments of the present application relate to the technical field of intelligent cabins, and particularly relate to a human-factor intelligent driving behavior prediction method and system, and a terminal device and a storage medium. The method comprises: acquiring a physiological signal of a driver; performing a fast Fourier transform on the physiological signal to solve an amplitude-frequency characteristic; then, performing multi-cycle decomposition on the physiological signal; next, performing dimension raising on decomposed data; importing the data into a multi-modal synchronous data fusion layer in combination with vehicle road scene video frame prediction data, so as to obtain a corresponding multi-scale three-dimensional feature; performing analysis processing on the multi-scale three-dimensional feature by means of a three-dimensional backbone network layer, and outputting a corresponding target feature; and finally, performing text generation on the target feature by means of a driving behavior interpretation layer and a driving behavior reasoning layer, so as to obtain interpretation information and reasoning information for predicting a human-factor intelligent driving behavior. The human-factor intelligent driving behavior prediction method and system, and the terminal device and the storage medium which are provided in the present application can improve the effect of predicting a driving behavior.