Jürgensen M, Vossiek M, Michel JF (2024)
Publication Language: English
Publication Type: Conference contribution, Conference Contribution
Publication year: 2024
Event location: Lisboa, Portugal
Radar sensors play a major role in the environmental perception of ADAS [1]. Their robustness in adverse weather conditions particularly differentiates the radar from other sensors used in ADAS, such as camera and lidar [2]. Furthermore, radar sensors are cheaper than lidar sensors, making them feasible for a wide range of applications in vehicles. Due to the physical working principle, radar sensors can achieve accurate measurements in range and velocity. However, in ADAS applications, not only range and velocity measurements but also measurement of the angular position in both the azimuth and elevation domains are crucial. In future applications, radars ought to be able to resolve fine details in the environment, such as
Manuel Jürgensen, Martin Vossiek, and Juan Carlos Fuentes Michel
vulnerable road users in different situations. Increasing the resolution of radar applications is a field of current research in the ADAS field [3].
Multiple papers have shown that deep learning has great potential in the domain of radar signal processing. In particular, in the case of angular processing, deep neural networks (DNNs) show promising results [4], [5]. DNNs can increase the resolution of different radar sensors using the same hardware and antenna layouts [4]. However, training DNNs relies on labelled data availability. Current datasets on radar data cover only some signal processing steps and sometimes lack annotations [6].
Furthermore, generic radar datasets that can be applicable to different radar systems cannot be retrieved through measurements. This is due to the nature of the radar cube and the information stored within it. Zhou et al. [6] reviewed different radar datasets for diverse types of radar. From an industrial perspective, developing a DNN for detection purposes is reasonable only if the DNN fits the sensor in use. This creates a problem with data collection and labelling. For every new sensor in development, we need to create a new dataset and a new set of labels, which is resource-intensive and time-consuming. A solution to this problem is proposed with a ray tracing simulation based on synthetic environments, in which Schuessler et al. [7] created realistic results in their radar data simulations. However, this approach requires creating artificial scenarios or using pregenerated digital worlds, each presenting its challenges related to realism. Ensuring realistic scenarios and a broad mix of object and scenario types is critical, as is addressing the challenges of simulating radar cross-section (RCS) [6].
In this paper, we present a method for generating datasets that are partly synthetic and partly data-driven. We use recorded detections from a production high-resolution (HR) radar sensor and accumulate them over consecutive radar cycles to increase the resolution. For our raw radar spectrum simulation, we treat each accumulated detection as a scatter centre and generate a sine wave in the intermediate frequency (IF) domain. Our signal processing framework then processes the raw spectrum, computes range and velocity fast Fourier transform (FFT), performs beamforming, and determines the local maxima through a constant false alarm rate (CFAR) algorithm. In this way, we generate every step of the radar data cube that is of interest for radar performance analysis. Our framework’s advantages lie in the practicability and versatility of the approach. According to our findings, our approach has proven to be economical in the predevelopment of radar and in generating data for possible use in deep learning applications.
APA:
Jürgensen, M., Vossiek, M., & Michel, J.F. (2024). A PRACTICAL DATA GENERATION FRAMEWORK FOR 4D AUTOMOTIVE MIMO RADAR: ENABLING DEEP LEARNING AND RADAR PERFORMANCE ANALYSIS. In IEEE (Eds.), Proceedings of the The 9th European Congress on Computational Methods in Applied Sciences and Engineering ECCOMAS Congress 2024. Lisboa, Portugal.
MLA:
Jürgensen, Manuel, Martin Vossiek, and J.C. Fuentes Michel. "A PRACTICAL DATA GENERATION FRAMEWORK FOR 4D AUTOMOTIVE MIMO RADAR: ENABLING DEEP LEARNING AND RADAR PERFORMANCE ANALYSIS." Proceedings of the The 9th European Congress on Computational Methods in Applied Sciences and Engineering ECCOMAS Congress 2024, Lisboa, Portugal Ed. IEEE, 2024.
BibTeX: Download