Movement Mode Learning with Clustered Pedestrian Trajectory Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing model parameter learning methods for movement mode models in environments with multiple moving objects are prone to bias, leading to inefficient learning due to biased data acquisition, which can be exacerbated by unnecessary data collection.
Innovation Solution
A method that involves acquiring link data of synchronized movement trajectories, extracting condition satisfying data, clustering feature amount vectors, and learning model parameters using a machine learning algorithm to ensure unbiased data for efficient model parameter learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If learning data is acquired by directly using movement trajectories of reference pedestrian and surrounding pedestrians, then data acquisition is simple, but the learning data becomes biased depending on acquisition situation and environment
Solution Approach 1:
The patent applies preliminary action by performing clustering analysis on movement trajectories before using them as learning data. The system clusters surrounding pedestrians into multiple groups based on their movement patterns, and then selects representative trajectories from each cluster. This preprocessing step ensures that the learning data is representative and unbiased before the actual model training begins.
Solution Approach 2:
The patent introduces clustering analysis as an intermediary process between raw trajectory acquisition and model training. By using clustering algorithms to group and select representative trajectories, the system creates an intermediate layer that filters out biases from the raw data, providing more balanced learning data to the model.
2Measurement precision
If a large number of learning data are acquired to prevent bias, then bias occurrence is prevented, but data acquisition efficiency deteriorates due to collecting more than necessary
Solution Approach 1:
The patent extracts only the necessary and representative learning data by using clustering analysis. Instead of collecting all available trajectory data, the system identifies and extracts representative trajectories from each cluster, removing redundant and biased data while maintaining data representativeness. This extraction process improves efficiency by collecting only what is necessary.
Solution Approach 2:
The patent changes the parameter of data selection from random or exhaustive collection to cluster-based representative selection. By transforming the selection criterion from quantity-focused to quality-focused (representativeness), the system achieves unbiased learning data with fewer samples, thereby improving acquisition efficiency.
3Device complexity
If movement trajectories are acquired without clustering analysis, then data processing is simple, but the model parameter cannot be appropriately learned due to biased learning data
Solution Approach 1:
The patent performs clustering analysis as a preliminary step before model training to ensure data quality. By pre-processing the trajectory data through clustering and selecting representative samples, the system guarantees that the learning data is representative and unbiased, which is essential for appropriate model parameter learning.
Solution Approach 2:
The patent introduces clustering analysis as an intermediary process that bridges raw trajectory data and model training. This intermediate processing step ensures data representativeness without requiring overly complex processing, maintaining a balance between processing complexity and learning reliability.
Data Source
AI summary
A learning device 30 extracts feature amount vectors from link data in which movement trajectories of a reference pedestrian M1 and a nearby pedestrian M2 are linked, links the feature amount vectors to condition satisfying data, determines the number of clusters, clusters the feature amount vectors of all clusters corresponding to the number of clusters, acquires the condition satisfying data to which the feature amount vectors are linked as learning data when the clustering of the feature amount vectors of all clusters corresponding to the number of clusters has converged, and learns a model parameter of a movement mode model by a predetermined machine learning algorithm using the learning data.