Abstract
In this paper, we discuss the usage of deep Convolutional Autoencoders (CAE) for denoising Range-Doppler (RD) maps of an FMCW radar in a near-field situation with pedestrians and cyclists as moving objects. Traditional methods for noise reduction such as CFAR with Peak Detection (PD) have poor performance under highly noisy environments, especially when objects have low reflectivity, like pedestrians. We propose the use of Convolutional Autoencoders, in various configurations, to overcome those limitations. Due to its Artificial Neural Network nature, CAE can extract features, learn to identify patterns and recognize moving objects - such as humans and bicycles - while the traditional method acts passively. The results indicate that the usage of CAE overcame CFAR with PD by 76.8%, on average, for reconstructing objects in their correct positions. The results of this work may have many applications, for instance, detecting distant moving objects that are too faint for regular detection.
Original language | English |
---|---|
Title of host publication | 2020 IEEE International Radar Conference (RADAR) |
Publisher | IEEE |
Pages | 630-635 |
Number of pages | 6 |
ISBN (Electronic) | 978-1-7281-6813-5 |
DOIs | |
Publication status | Published - 11 Jun 2020 |
Event | IEEE Radar Conference 2020 - Virtual Conference Duration: 21 Sep 2020 → 25 Sep 2020 https://www.radarconf20.org/ |
Conference
Conference | IEEE Radar Conference 2020 |
---|---|
Abbreviated title | RadarConf 2020 |
City | Virtual Conference |
Period | 21/09/20 → 25/09/20 |
Internet address |
Keywords
- Autoencoder
- Convolutional
- Deep Learning
- Doppler-Range
- Radar
- 22/2 OA procedure