Radar Denoising Autoencoder False Alarm Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current radar processing systems face challenges in effectively discriminating between real targets and false alarms, particularly for low Radar Equivalent Surface (SER) targets, as existing intermediate processing methods either fail to utilize kinematic information or incur high computational costs, especially in dense environments with maneuvering targets.
Innovation Solution
Implementing an intermediate denoising process using a deep learning algorithm, specifically a convolutional autoencoder, to differentiate between real targets and false alarms by training on real radar data and applying it in real-time to radar signals, thereby reducing false alarms and enhancing detection performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If turn integration (TIT) is used to filter detections based on temporal coherence, then the number of detections entering tracking is reduced and computational cost is lowered, but kinematic information is not utilized and integration time must be reduced
Solution Approach 1:
The patent merges temporal coherence filtering (TIT) with kinematic coherence filtering (TBD) into a unified denoising process. The convolutional autoencoder simultaneously processes detection data across multiple time scans and evaluates kinematic consistency, combining the advantages of both approaches without their individual limitations.
Solution Approach 2:
The denoising unit serves multiple functions: it filters false alarms based on temporal coherence, evaluates kinematic consistency, and outputs cleaned detection data for tracking. This multi-functional approach replaces the need for separate TIT and TBD processing chains.
2Reliability
If Track Before Detect (TBD) is used to exploit kinematic coherence, then filtering over longer time horizon is achieved, but computational cost becomes high and a priori motion modeling is required
Solution Approach 1:
The patent replaces the traditional mechanical/mathematical motion modeling approach of TBD with a data-driven deep learning approach. The convolutional autoencoder learns kinematic patterns directly from radar data without requiring a priori motion models, eliminating the complexity of manual motion model selection and parameter tuning.
Solution Approach 2:
The patent changes the approach from fixed a priori motion models to adaptive parameter learning. The denoising unit's parameters (weights and biases of the convolutional autoencoder) are trained on historical radar data to automatically adapt to different target motion patterns, eliminating the need for manual motion model configuration.
3Measurement precision
If intermediate processing is implemented to lower detection threshold, then discrimination between false alarms and real targets is improved, but false alarm rate remains high in dense environments
Solution Approach 1:
The patent applies preliminary denoising processing to detection data before it enters the tracking unit. The convolutional autoencoder pre-filters false alarms by learning the statistical characteristics of both true targets and false alarms from training data, reducing the false alarm burden on downstream tracking processes.
Solution Approach 2:
The denoising unit acts as an intermediary between the detection processing unit and the tracking processing unit. It receives raw detection data, applies learned denoising transformations, and outputs cleaned detection data, mediating the information flow between detection and tracking while removing false alarms.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
This system comprises: a detection unit (12), to produce, from a radar signal, a detection map; an intermediate unit, to produce, from the detection map, a processed detection map; and a tracking unit (14), to produce, from the processed detection map, tracks.This system is characterized in that the intermediate unit is a denoising unit (20) suitable for processing a detection map at a current time (C(t)) to obtain a denoised detection map at the current time (CN(t)), the denoising unit (20) comprising: a data structuring module (22), suitable for associating an input tensor (TE(t)) with the detection map at the current time and several detection maps at times prior to the current time; a convolutional autoencoder (24), suitable for identifying, in the input tensor, signatures characteristic of real targets, and for producing at output a denoised detection map (M(t)); and a cleaning module (26) filtering the detection map at the current time with the denoised detection map to obtain the denoised detection map at the current time.