University of Twente Student Theses
Exploring Optimization and Cross-Modal Learning for Radar Target Detection in Range-Doppler Maps
Mahovic, Borna (2025) Exploring Optimization and Cross-Modal Learning for Radar Target Detection in Range-Doppler Maps.
Full text not available from this repository.
| Full Text Status: | Access to this publication is restricted |
| Embargo date: | 31 March 2027 |
| Abstract: | This work investigates the limitations of traditional Constant False Alarm Rate (CFAR) detectors in automotive radar systems and proposes machine learning based methods to enhance target detection performance and eliminate the need for ground truth labels. Traditional CFAR algorithms, constrained by fixed hyperparameters, often suffer from target masking, ghost targets, and sensitivity trade-offs, particularly in cluttered and dynamic environments. To address these challenges, firstly hyperparameter optimization (HPO) is performed to explore the full potential of Smallest-Of CFAR(SO-CFAR), Greatest-Of CFAR (GO-CFAR), and Ordered-Statistic CFAR (OS-CFAR) algorithms. The resulting configurations reveal a trade-off between detection sensitivity and false alarm rates, with optimized settings significantly increasing sensitivity compared to baseline performance. Building on this, a novel neural network (NN) detector is introduced that leverages Light Detection and Ranging (LiDAR) point clouds as precise ground truths in a cross-modal training framework. The NN detector, inspired by U-Net architecture is trained using a custom loss function based on a weighted Chamfer distance metric. This loss function incorporates balancing and profile matching terms to guide the network toward generating dense, accurate confidence maps from raw range-Doppler inputs. Experiments on both static radar and sensor-car datasets demonstrate that the proposed NN detector outperforms traditional CFAR methods in sensitivity while maintaining competitive false alarm rates, especially for larger, moving targets. These findings highlight the potential of integrating HPO and cross-modal deep learning to overcome the inherent limitations of conventional radar processing techniques in automotive applications. |
| Item Type: | Essay (Master) |
| Faculty: | EEMCS: Electrical Engineering, Mathematics and Computer Science |
| Subject: | 54 computer science |
| Programme: | Embedded Systems MSc (60331) |
| Link to this item: | https://purl.utwente.nl/essays/106080 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page