Modality-Independent Feature Extraction for Multi-Sensor Object Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for object classification in driverless navigation require separate training of classification modules for each sensor modality, which is time-consuming and error-prone, especially for lidar and radar data, due to the difficulty in annotating and the scarcity of annotated data sets compared to image data.
Innovation Solution
A method that extracts modality-independent features from measuring data, allowing a single classification unit to be trained across different sensor modalities, such as lidar, radar, and image, enabling the reconstruction of data from one modality using features from another, thereby eliminating the need for separate training for each sensor modality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If separate classification modules are trained for each sensor modality, then classification accuracy for each modality is improved, but training time and effort increase significantly
Solution Approach 1:
The patent applies universality by training a single classification module that can process multiple sensor modalities (lidar, radar, image sensors) simultaneously. The classification unit is designed to accept feature vectors from different sensor types and perform classification across all modalities, eliminating the need for separate training processes for each sensor while maintaining classification accuracy.
2Measurement precision
If separate classification modules are trained for each sensor modality, then modality-specific classification performance is improved, but the complexity of the system increases
Solution Approach 1:
The patent merges multiple separate classification modules into a single unified classification unit. Instead of having distinct classifiers for lidar, radar, and image data, the system combines them into one classification module that processes feature vectors from all sensor modalities, thereby reducing system complexity while maintaining classification performance.
3Measurement precision
If annotated data sets are used for training, then classification accuracy is improved, but the annotation process becomes time-consuming and error-prone
Solution Approach 1:
The patent applies universality by enabling a single classification module to be trained on annotated data from one sensor modality (particularly image sensors which have abundant annotated data) and then applied to multiple sensor modalities. This eliminates the need to create separate annotated data sets for lidar and radar, significantly reducing annotation effort while maintaining classification accuracy across all modalities.
Data Source
AI summary
A method for classifying objects which comprises a provision of measuring data from a sensor for a feature extraction unit as well as extraction of modality-independent features from the measuring data by means of the feature extraction unit, wherein the modality-independent features are independent of a sensor modality of the sensor, so that a conclusion to the sensor modality of the sensor is not possible from the modality-independent features.