Resolution-Adaptive Models for Movable Object Direction Evaluation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for predicting the moving direction of movable objects, such as pedestrian crossing intention, do not necessarily improve accuracy with an increase in the amount of information used, as the quality of that information can vary and affect the prediction accuracy.
Innovation Solution
An evaluation method and system that utilize a plurality of trained models, including a simplified model for low resolution images and a detailed model for higher resolution images, to enhance the accuracy of moving direction evaluation by selecting the appropriate model based on image resolution, using whole and part information accordingly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If various types of information (vehicle speed, pedestrian bounding box, pose key points, local context, global context) are used for prediction, then prediction accuracy of pedestrian crossing intention is improved, but the complexity of the evaluation system increases
Solution Approach 1:
The system segments the evaluation process into two distinct models: a simplified model that processes only whole information (global context) and a detailed model that processes both whole information and part information (local context, pose key points). This segmentation allows the system to handle different information types through specialized models, improving prediction accuracy while managing complexity through modular architecture.
Solution Approach 2:
The system applies local quality by having the detailed model process specific local features (pose key points, local context) only when necessary, while the simplified model handles global context universally. This allows different parts of the system to have different processing capabilities matched to the requirements of the input data, optimizing both accuracy and complexity management.
2Measurement precision
If more information is used for evaluation, then accuracy of evaluation may be improved, but the amount of processing required increases
Solution Approach 1:
The system dynamically selects between the simplified model and the detailed model based on the characteristics of the input data. When local context and pose information are available and beneficial, the detailed model is used; otherwise, the simplified model is selected. This dynamic adaptation optimizes processing efficiency while maintaining evaluation accuracy.
Solution Approach 2:
The system changes the processing parameters by switching between different model configurations based on data availability. The simplified model uses fewer processing parameters (only whole information), while the detailed model uses additional parameters (part information). This parameter adaptation allows the system to balance accuracy requirements with processing efficiency.
3Device complexity
If a single model processes all types of information, then system complexity is reduced, but accuracy of evaluation deteriorates
Solution Approach 1:
The system segments the evaluation function into two specialized models: one optimized for whole information processing and another optimized for both whole and part information processing. This segmentation ensures that each model can be tuned for its specific information type, maintaining high accuracy while the modular structure manages complexity through clear functional separation.
Solution Approach 2:
The detailed model serves a universal function by being capable of processing both whole information and part information, while the simplified model handles whole information universally. This multi-functionality allows the system to adapt to different input scenarios without requiring entirely separate specialized models for each case.
Data Source
AI summary
The evaluation method is performed by an arithmetic circuit accessible to a storage device storing: a simplified model trained to, in response to input of whole information based on a whole of a movable object in image information, output evaluation of a moving direction of the object; and a detailed model trained to, in response to input of the whole information and part information based on part(s) of the object, output the evaluation. The evaluation method includes, when a resolution of an image of the object detected from target image information is smaller than a threshold value, inputting the whole information into the simplified model to output the evaluation. The evaluation method includes; when the resolution is equal to or greater than the threshold value, inputting the whole information and the part information into the detailed model to output the evaluation.