Adaptive Radar Filtering for Reliable Multi-Target Tracking

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing radar systems face challenges in reliably tracking multiple targets, particularly when the scene dynamics deviate from the trained environment, leading to potentially erroneous tracking results.

Innovation Solution

A radar device equipped with a machine learning logic that includes a policy network to set digital filter parameters and a reward value generating network to detect out-of-distribution scenes, ensuring reliable tracking by analyzing the distribution of reward values generated by multiple heads.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If digital filtering is used to track multiple targets, then tracking capability is improved, but reliability deteriorates when scene dynamics deviate from trained environment

Engineering Contradiction:
Improvetracking capabilityVSAvoidtracking reliability
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The system uses a reward value generating network that provides feedback signals to the policy network. This feedback mechanism allows the system to learn from previous performance and adjust its filtering parameters dynamically, improving reliability when scene dynamics deviate from trained environments by continuously adapting to new conditions based on reward signals.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The policy network dynamically adjusts the parameters of the digital filter based on the current radar scene and learned policies. By changing filter parameters adaptively rather than using fixed parameters, the system maintains high tracking reliability across diverse and changing environmental conditions while preserving the enhanced tracking capability for multiple targets.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If machine learning logic is added to set filter parameters, then tracking accuracy is improved, but device complexity increases

Engineering Contradiction:
Improvetracking accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The machine learning components (policy network and reward value generating network) are integrated into the existing radar processing pipeline. The ML logic combines with the digital filter and signal processing circuits in a unified architecture, improving tracking accuracy while minimizing the increase in device complexity through functional integration rather than adding separate complex subsystems.

Inventive Principle:
Principle #5Merging (Combining)

3Reliability

If multiple heads are used in reward value generating network, then detection of out-of-distribution scenes is improved, but computational load increases

Engineering Contradiction:
Improveout-of-distribution detectionVSAvoidcomputational load
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The reward value generating network is divided into multiple independent heads, each responsible for evaluating specific aspects of the radar scene. This segmentation allows parallel processing of different detection tasks, improving out-of-distribution scene detection reliability while managing computational load through distributed processing architecture.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4300132B1Radar device and method of operating a radar device
Publication Date: 2025.08.27 INFINEON TECHNOLOGIES AG
  • EP4300132B1 patent drawingFigure 1~2
  • EP4300132B1 patent drawingFigure 3
  • EP4300132B1 patent drawingFigure 4

AI summary

A radar device is provided, comprising a radar frontend (11) configured to send radar signals and to receive reflected radar signals, processing circuitry (12) configured to provide digital radar data (rd, rd2) based on the received reflected radar signals, and a digital filter (13) configured to process the digital radar data (rd, rd2) to obtain information about objects which reflected the radar signals. The device further comprises a machine learning logic (14) with a policy network configured to set the parameters of the digital filter based on the digital radar data (rd, rd2), and a reward value generating network including a plurality of heads, each head configured to provide a respective expected reward value for a setting of parameters by the policy network. The radar device is further configured to detect that a scene (10) captured by the radar device is not reliably processable based on a distribution of the expected reward values generated by the plurality of heads.