Radar Object Classifier Trained via Image Sensor Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object classification systems require large amounts of labelled data, which can be resource-intensive and often lack the specific data types or formats needed for training, particularly for non-image data like radar, leading to inefficiencies in training neural networks for accurate object detection.
Innovation Solution
A method that utilizes both image and radar sensors to generate classifications, where the image-based object classifier trains the radar-based object classifier using image data when specific conditions are met, allowing for continuous improvement and adaptation of the radar-based classifier's performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If large amounts of labelled data are used to train object classifiers, then classification accuracy is improved, but resource consumption and data generation cost increase significantly
Solution Approach 1:
The system performs preliminary action by using the image-based classifier (trained on readily available image data) to generate preliminary classifications that serve as training labels for the radar-based classifier. This preliminary classification from the image sensor provides the ground truth needed to train the radar classifier without requiring manual labelling of radar data, thus reducing the quantity of labelled radar data needed while maintaining classification accuracy.
2Loss of time
If existing labelled data sets are used for training, then training time is reduced, but data format compatibility and relevance to specific object classes decrease
Solution Approach 1:
The system introduces an intermediary approach by using the image-based classifier as a bridge between readily available image data and the radar classification task. The image-based classifier processes easily obtainable image data to generate classifications that serve as training labels for the radar-based classifier, thus enabling training without requiring pre-collected labelled radar datasets while maintaining adaptability to specific object classes of interest.
3Measurement precision
If manual labelling of training data is performed, then data quality and relevance to specific object classes are improved, but time consumption and cost increase
Solution Approach 1:
The system implements self-service by having the image-based classifier automatically generate training labels for the radar-based classifier without human intervention. The image sensor continuously captures data, the image-based classifier processes this data to generate classifications, and these classifications are automatically used to train the radar-based classifier, eliminating the need for manual labelling while maintaining high data quality and relevance to specific object classes.
4Reliability
If radar-based object classification is used independently, then detection capability in challenging visual conditions is improved, but classification accuracy decreases due to lack of training data
Solution Approach 1:
The system merges the strengths of both image-based and radar-based classification by combining their training approaches. The image-based classifier is trained on readily available image data, and its classifications are used to train the radar-based classifier. This merging allows the radar-based classifier to achieve high classification accuracy by learning from the image-based classifier's outputs, while maintaining its ability to operate independently in challenging visual conditions where radar detection is more reliable.
Data Source
AI summary
Apparatus are provided for generating an object classification for an object, the apparatus comprising an image sensor, a radar sensor, and a processing unit, the processing unit configured to perform the steps of: receiving image data for the object from the image sensor, operating an image-based object classifier on the image data to generate an image-based object classification, receiving radar data for the object from the radar sensor, operating a radar-based object classifier on the radar data to generate a radar-based object classification, selecting between the image-based object classification and the radar-based object classification to output as the object classification for the object, determining if a training condition is met by the radar-based object classification, and training the radar-based object classifier using the image-based object classification when the training condition is met by the radar-based object classification.