CNN Radar Signal Processing for FMCW Target Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing radar signal processing techniques, such as those using CFAR detection, face challenges in setting appropriate detection thresholds, leading to detection failures or overdetection due to varying environmental conditions, which affects the reliability of target detection in FMCW radar systems.
Innovation Solution
The use of a convolutional neural network (CNN) is introduced to process two-dimensional data obtained from Fourier transforms of FMCW radar signals, where windowing is applied to reduce CNN size, and probabilities of target existence are calculated, allowing for improved target identification by selecting range indices based on these probabilities, followed by a second Fourier transform to determine speed-index-specific coefficient values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If CFAR detection with variable detection threshold is used to maintain constant false alarm rate, then detection reliability is improved, but device complexity and difficulty of operation increase due to the need to select appropriate threshold according to various environmental conditions
Solution Approach 1:
The system performs self-service by automatically selecting the detection threshold through CNN-based target probability estimation. Instead of requiring manual threshold adjustment for different environmental conditions, the system autonomously adapts by processing radar signals through the neural network, which learns optimal threshold settings from training data representing various environmental scenarios.
Solution Approach 2:
The patent replaces the mechanical/manual threshold setting process with an intelligent system based on convolutional neural networks. The CNN model substitutes the traditional manual or rule-based threshold selection mechanism, using learned patterns from training data to automatically determine appropriate thresholds for different environmental conditions, thereby eliminating the complexity of manual threshold adjustment.
2Measurement precision
If CFAR detection with multiple processing steps is used to remove noise and clutter, then detection precision is improved, but productivity decreases due to the multi-step processing requirement
Solution Approach 1:
The patent merges multiple sequential processing steps (FFT, windowing, CFAR detection, target selection) into a single integrated CNN processing pipeline. The convolutional neural network performs all these functions simultaneously through its layered architecture, where feature extraction, noise filtering, and target detection occur in parallel within the network, eliminating the sequential bottlenecks of traditional multi-step processing.
Solution Approach 2:
The system transitions from traditional one-dimensional signal processing to two-dimensional data processing by applying Fourier transforms at every pulse repetition interval and organizing results in a 2D matrix. This dimensional change enables the CNN to process range-Doppler information simultaneously, improving both detection precision and processing efficiency by exploiting spatial correlations in the 2D data structure.
3Reliability
If full two-dimensional data is processed by CNN without windowing, then detection reliability is improved, but device complexity and computational load increase
Solution Approach 1:
The patent applies segmentation by dividing the full two-dimensional radar data into smaller sub-regions or windows before feeding them to the CNN. This segmentation reduces the input data size for each CNN processing operation, thereby reducing the required model size and computational resources while maintaining detection reliability through systematic coverage of the entire data space through multiple processed windows.
Solution Approach 2:
The system uses partial action by processing only selected regions of the two-dimensional data through the CNN rather than the entire dataset at once. Windowing allows the system to focus computational resources on specific regions of interest in the range-Doppler map, reducing the effective input size to the CNN while still achieving comprehensive target detection through systematic scanning of different data regions.
Data Source
AI summary
Two-dimensional data obtained by performing a Fourier transform on a digitally converted FMCW radar signal at every pulse repetition interval over N pulse repetition intervals is input to a convolutional neural network (CNN) to find the probabilities of the existence of a target in range indices. The range indices, i.e., bit frequencies are selected based on the probabilities of the existence of the target. In order to reduce the size of the CNN, windowing is applied to the two-dimensional data. A speed-index-specific coefficient value may be calculated by re-performing a Fourier transform on range data for the selected bit frequencies. Thus, the range and speed of the target may be calculated.