Radar Semantic Segmentation Using Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing radar systems discard a majority of the information from range-Doppler maps, primarily using only signal/noise values for processing, and fail to utilize spatial distribution and antenna characteristics, limiting the derivation of global scenarios from radar recordings.
Innovation Solution
An evaluation device employing an artificial neural network for semantic segmentation of radar recordings, which processes radar data including distances, angles, and interference signals to recognize structures and form relationships with objects, using forward and backward propagation to refine characteristics and weighting factors for neural connections.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If threshold values are set high to discard clutter, then false alarm rate is reduced, but information loss increases
Solution Approach 1:
The patent segments the radar data processing into multiple stages: initial clutter filtering using CFAR, followed by semantic segmentation using neural networks that process both filtered and discarded data. This allows progressive refinement where early filtering reduces false alarms, but later processing recovers useful information from previously discarded data.
Solution Approach 2:
The patent changes the processing parameters dynamically by adjusting threshold values adaptively and applying different processing strategies to different data segments. The neural network learns optimal parameter settings from training data, allowing the system to maintain low false alarm rates while preserving relevant information through adaptive parameter adjustment.
2Device complexity
If only signal/noise values are used for processing, then computational complexity is reduced, but measurement precision deteriorates
Solution Approach 1:
The patent transitions from one-dimensional signal amplitude processing to multi-dimensional processing by incorporating spatial distribution coordinates, antenna characteristics, and temporal information into the neural network input. This dimensional expansion enables precise object recognition while the network's efficient architecture manages the computational complexity of processing these additional dimensions.
Solution Approach 2:
The patent performs preliminary processing steps including CFAR filtering, range-Doppler map generation, and data formatting before neural network processing. These preliminary actions organize and pre-process the raw radar data into structured formats, reducing the computational burden on the neural network while preserving all necessary information for precise object recognition.
3Productivity
If local maxima are selected for processing, then processing speed is improved, but loss of information increases
Solution Approach 1:
The patent segments the target selection process into two phases: initial rapid selection of local maxima for quick processing, followed by comprehensive neural network analysis that processes both selected maxima and previously discarded data. This segmentation enables fast initial processing while ensuring complete information utilization in the final analysis stage.
Solution Approach 2:
The neural network acts as an intermediary that reconciles the conflict between fast maximum-based processing and comprehensive information analysis. It takes both the quickly identified local maxima and the previously discarded data as inputs, integrating them to produce accurate object recognition without requiring exhaustive processing of all raw data points.
4Reliability
If adaptive threshold adjustment is used, then reliability is improved, but device complexity increases
Solution Approach 1:
The neural network performs self-adjustment of processing parameters by learning optimal threshold values and processing strategies from training data. This self-service capability allows the system to adapt to different environmental conditions and clutter types automatically, improving reliability while the learned parameters reduce the need for complex real-time adjustment mechanisms.
Data Source
AI summary
An evaluation device for obtaining a segmentation of an environment from a radar recording of the environment, that has an input interface configured to obtain initial training data, where the initial training data comprise radar data of the radar recording and initial characteristics of objects located in the environment recorded with a radar sensor that generates the radar recordings, and where the evaluation device is configured to forward propagate an artificial neural network with the initial training data to obtain second characteristics of the objects determined with the artificial neural network in the forward propagation, and to obtain weighting factors for neural connections of the artificial neural network through backward propagation of the artificial neural network with the differences between the second characteristics and the initial characteristics, in order to obtain the segmentation of the environment through renewed forward propagation with these radar data.