Autoencoder-Based Target Detection in Automotive MIMO FMCW Radar System.

Area of Science:

• Radar Systems Engineering
• Artificial Intelligence in Automotive Technology
• Signal Processing

Background:

• Constant False Alarm Rate (CFAR) algorithms are standard for target detection in automotive Frequency-Modulated Continuous Wave (FMCW) radar.
• Improper parameter settings in CFAR algorithms can significantly degrade target detection performance.

Purpose of the Study:

• To propose and evaluate a convolutional neural network-based autoencoder (AE) as an alternative to CFAR algorithms for target detection in Multiple-Input Multiple-Output (MIMO) FMCW radar systems.
• To optimize the AE architecture for improved signal-to-noise ratio in target detection results.

Main Methods:

• Developed a convolutional neural network-based autoencoder (AE) to replace conventional CFAR algorithms.
• The AE compresses detection results at the encoder and recovers significant signal components at the decoder.
• Systematically varied the number of hidden layers and filters within the AE to determine optimal structure for high signal-to-noise ratio.

Main Results:

• The proposed AE-based target detection method demonstrated superior performance compared to conventional CFAR algorithms.
• Correlation coefficients were calculated between AE detection results and actual target positions.
• The AE-based method achieved a high similarity with a correlation coefficient of 0.73 or higher.

Conclusions:

• The convolutional neural network autoencoder offers a promising approach for enhancing target detection in automotive MIMO FMCW radar systems.
• The AE method provides a more robust and accurate alternative to traditional CFAR algorithms, especially when CFAR parameters are sub-optimal.
• Optimizing AE architecture is crucial for achieving high signal-to-noise ratios and improved detection accuracy.