Ultrasonic Sensor Training Data Balancing for Object Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Ultrasonic sensor systems face challenges in providing homogeneous training datasets for object classification models due to varying densities of data collection features with respect to distance, leading to over-representation in certain distance ranges and inadequate data in others, affecting the reliability and complexity of the trained models.
Innovation Solution
A method is introduced to select and normalize training datasets by considering the relative distance, velocity, and movement trajectory of surrounding objects, ensuring equal distribution across all collection situations, and weighting datasets based on velocity and age to compensate for data accumulation in specific distance ranges.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If training data is collected without considering distance-based density, then data collection is simple, but training dataset homogeneity deteriorates due to over-representation in certain distance ranges
Solution Approach 1:
The patent applies parameter changes by introducing distance-based weighting factors that modify the contribution of training samples according to their distance from the ultrasonic sensor. This transforms the uniform sampling approach into a distance-aware sampling strategy, where samples from different distance ranges are weighted differently to achieve homogeneous representation across all distance zones.
Solution Approach 2:
The patent implements local quality by applying different selection criteria and weighting factors to different distance ranges. Instead of treating all training samples uniformly, the system applies localized quality control measures specific to each distance zone, ensuring that each local region (near, mid, far) contributes appropriately to the overall training dataset homogeneity.
2Productivity
If all candidate training datasets are used, then data quantity increases, but model training efficiency decreases due to redundant data in over-represented distance ranges
Solution Approach 1:
The patent applies the extraction principle by selectively removing or down-weighting redundant training samples from distance ranges that are already well-represented. The system extracts only the necessary subset of training data required for homogeneous coverage, discarding excess samples that would otherwise waste computational resources during model training.
Solution Approach 2:
The patent implements partial action by collecting more candidate training datasets than ultimately needed, then applying filtering and weighting to retain only the appropriate portion. This excessive initial collection ensures sufficient coverage of all distance ranges, followed by selective retention based on distance-based weighting to achieve optimal training efficiency.
3Reliability
If velocity and age weighting is not applied, then data processing is simpler, but training dataset balance deteriorates due to data accumulation in specific distance ranges
Solution Approach 1:
The patent applies feedback by continuously monitoring the distribution of training samples across different distance ranges and adjusting the weighting factors accordingly. The system uses the observed data accumulation patterns as feedback to modify future sampling weights, creating a self-regulating mechanism that maintains training dataset balance despite varying data collection conditions.
Solution Approach 2:
The patent implements dynamics by making the weighting factors adaptive rather than static. The weighting mechanism dynamically adjusts based on the current state of data accumulation in different distance ranges, allowing the system to respond to changing conditions during data collection and maintain balance throughout the training dataset assembly process.
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
This approach ensures reliable and efficient training of object classification models by providing balanced training datasets, preventing over-representation in certain distance ranges and enhancing feature quality, allowing for accurate classification of object properties like traversability and collision relevance.
Implementation Method 1
a series of ultrasonic transducers that emit ultrasonic signals to receive and evaluate ultrasonic signals reflected from surrounding objects
Data Source
AI summary
A method for providing training datasets for training an object classification model for object classification in an ultrasonic sensor system is disclosed. The method includes (i) providing one or multiple survey scenarios in which at least one surrounding object within a collection range of the ultrasonic sensor system is moved along a trajectory relative to the ultrasonic sensor system, (ii) collecting the ultrasonic signals reflected at the surrounding object at chronologically successive collection situations and respective identification of collection features depending on reflected ultrasonic signals collected during a respective collection situation, (iii) determining a candidate training dataset for each collection situation by associating a classification vector specified by the survey situation, the elements of which each indicate an object property of at least one surrounding object, with the collection features, and (iv) considering the candidate training dataset of each of the collection situations as a training dataset depending on the relative distance from the at least one surrounding object from the ultrasonic sensor system and the relative distances of the surrounding object from the ultrasonic sensor system during previously measured collection situations of candidate training datasets determined.