Radar Image Augmentation for ML Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing radar systems face challenges in achieving high accuracy in object detection due to the limited richness of training datasets, particularly in generating sufficient variations and numbers of radar images for machine learning models.
Innovation Solution
An information processing apparatus that generates first and second radar images by executing signal processes based on different conditions, using these images as training data for a machine learning model to improve object detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the number of training data (radar images) is increased to improve detection accuracy, then the detection accuracy is improved, but the complexity of data preparation and processing increases
Solution Approach 1:
The patent applies preliminary action by pre-processing radar images through augmentation techniques (rotation, flipping, scaling, cropping) before they are used for training machine learning models. This prepares diverse training data in advance, allowing the model to learn from varied scenarios without requiring extensive manual data collection during the training phase.
Solution Approach 2:
The patent uses copying by generating multiple transformed versions of existing radar images through geometric transformations (rotation, flipping, scaling) and adding synthetic clutter. Instead of collecting new physical radar images, the system creates copies with modified properties to expand the training dataset diversity.
2Adaptability or versatility
If the variability of training data is increased to improve model robustness, then the model's generalization capability is improved, but the complexity of data generation increases
Solution Approach 1:
The patent applies dynamics by implementing dynamic data augmentation where training images are transformed with varying parameters (different rotation angles, flip combinations, scaling factors) during each training epoch or batch. This creates continuously varying training data that improves model robustness without requiring a static, excessively large dataset.
Solution Approach 2:
The patent uses parameter changes by modifying image properties such as rotation angle, flip direction, scaling ratio, and clutter addition intensity. These parameter variations generate diverse training samples from a limited set of base images, improving model adaptability while keeping the base dataset size manageable.
3Quantity of substance
If more signal processing conditions are applied to generate diverse radar images, then the richness of training data is improved, but the processing time and computational load increase
Solution Approach 1:
The patent applies segmentation by dividing the data preparation process into distinct stages: base image generation from radar signals, geometric transformations (rotation, flipping, scaling), clutter addition, and final image composition. This modular approach allows efficient processing of each stage independently and enables parallel processing of multiple transformation operations.
Solution Approach 2:
The patent uses partial action by selectively applying different augmentation techniques to different training samples based on their content and the specific training goals. Not all images undergo all transformations; instead, the system applies appropriate transformations selectively to achieve sufficient diversity without the computational cost of processing every image through every possible transformation.
Data Source
AI summary
According to one embodiment, an information processing apparatus includes a processor. The processor is configured to acquire a first observation signal from a radar device with a plurality of antennas configured to transmit a radar signal and to receive a radar echo based on a reflected wave of the radar signal, generate a first radar image by executing a signal process based on a predetermined first condition, and generate a second radar image by executing a signal process based on a second condition. The first and second radar images are used as training data for a learning model to detect an object to which the radar signal is transmitted.