Automotive frequency-modulated continuous wave (FMCW) radars, essential in Advanced Driver Assistance Systems, encounter mutual interference issues that degrade their detection capabilities. Model-based algorithms, though widely used, rely heavily on predetermined assumptions about the statistical properties. General-purpose black-box deep learning approaches, while effective in their training distribution, often lack flexibility and generalizability in dynamic environments. We introduce a novel hybrid method that combines model-based techniques with deep learning, treating interference mitigation as a source separation problem. Specifically, our method employs score-based deep generative networks to accurately capture the structure of FMCW interference. Additionally, we employ deep unfolding to accelerate inference, critical for automotive radar applications. Empirical results from simulated data demonstrate that the proposed algorithm outperforms the baseline models by 3.26 dB in signal-to-interference-plus-noise ratio in the presence of aggressive interference, and also shows good generalizability with measured data.

Radar technology plays a critical role in target detection, classification, and tracking. However, the computational demands of training deep neural networks (DNNs) on radar signals can be overwhelming, posing challenges for edge devices with limited energy and computing resources. In this article, we propose leveraging hyperdimensional computing (HDC), a brain-inspired computing paradigm, as an efficient alternative. HDC utilizes high-dimensional vectors for information representation and processing, offering robustness and energy efficiency. We propose a novel HDC classification algorithm named DynHD, with a dynamic HDC encoder that adapts to more challenging radar spectrum recognition tasks. We designed this mechanism to provide great flexibility to the HDC encoder that is otherwise fixed. Our evaluations demonstrate that HDC-based approaches achieve comparable accuracy to DNN-based methods with lower-computational complexity, making them suitable for resource-constrained devices. We achieve significant improvements in latency during training and inference phases, enabling efficient processing of radar signals on edge devices.

A neural network (NN) based multi-frame classification approach is proposed to solve the problem of classification of tracked objects. Initially, a baseline tracker is implemented that uses the classification output of an object detection network for classification. Afterwards, two approaches for multi-frame classification are applied to perform classification of tracked objects. The first approach aggregates points from multiple frames and applies a single frame NN for classification, whereas the second approach uses bidirectional long short term memory (BiLSTM) layers to process points from multiple frames. Extensive experiments on the opensource 2D RadarScenes dataset showed a consistent increase in track performance when using either of the two techniques for multi-frame classification.

This paper focuses on the challenge of estimating the 2D instantaneous ego-motion of vehicles equipped with an automotive radar. To further improve our previous study based on the weighted least squares (wLSQ) method and purpose-designed neural networks (NNs), this work proposes a new network architecture that supports local and global feature extraction as well as point-wise dynamic feature channel mixing. Compared with our previous work, the proposed method provides better estimation accuracy, lighter network size, and faster runtime performance.

Automotive radar has been extensively utilized in cars for many years as an essential sensor, primarily due to its robustness in extreme weather conditions, its capacity to measure Doppler information in the surrounding environment, and its cost-effectiveness. Recently, developments in radar technologies and the availability of open-source radar data sets have attracted more attention to radars and using them for perception tasks in deep learning based autonomous driving. However, annotating radar data for large-scale autonomous driving perception tasks is challenging, i.e., it is difficult for humans to label this data and often requires a semi-automatic approach that involves projecting labels from other sensors, such as cameras and LiDARs. The lack of high-quality labeled data has limited the performance of radar perception models. In this paper, we propose MVRAE, a Multi-View Radar AutoEncoder, which employs self-supervised learning to learn meaningful representations from multi-view radar data without any labels. Our approach is based on the intuition that a good representation for multi-view radar data, which includes range-angle, range-Doppler, and angle-Doppler views, should enable the reconstruction of one view solely from the representations of the other two views. Experimental results demonstrate that our proposed self-supervised method, that can be used as a pre-training step for autonomous driving task, allows the model to learn meaningful representations from unlabeled radar data and achieves enhanced label efficiency for downstream tasks, such as radar semantic segmentation. To the best of our knowledge, MVRAE is the first work that employs self-supervised learning and conducts systematic experiments with multi-view radar data.

A single frame radar-based multi-object tracker that aims to improve data association for better tracking performance is proposed. Firstly, a baseline tracker based on track-by-detection paradigm was implemented for automotive radar. Secondly, investigation on the performance of the tracker when tracking individual classes separately versus all classes together was performed. Thirdly, appearance features were extracted from a neural network and added as an additional metric to the cost matrix for improved data association. Extensive experiments on the 2D RadarScenes dataset and a 3D proprietary Lunewave dataset (in partnership with NXP Semiconductors) showed a consistent improvement in the tracking performance using the approach proposed by adding features extracted from a neural network.

Among various sensors for assisted and autonomous driving systems, automotive radar has been considered as a robust and low-cost solution even in adverse weather or lighting conditions. With the recent development of radar technologies and open-sourced annotated data sets, semantic segmentation with radar signals has become very promising. However, existing methods are either computationally expensive or discard significant amounts of valuable information from raw 3D radar signals by reducing them to 2D planes via averaging. In this work, we introduce ERASE-Net, an Efficient RAdar SEgmentation Network to segment the raw radar signals semantically. The core of our approach is the novel detectthen-segment method for raw radar signals. It first detects the center point of each object, then extracts a compact radar signal representation, and finally performs semantic segmentation. We show that our method can achieve superior performance on radar semantic segmentation task compared to the state-ofthe-art (SOTA) technique. Furthermore, our approach requires up to 20x less computational resources. Finally, we show that the proposed ERASE-Net can be compressed by 40% without significant loss in performance, significantly more than the SOTA network, which makes it a more promising candidate for practical automotive applications.

Radar is imperative for many automotive applications in detecting targets. Accurate direction of arrival (DOA) estimation is essential for maximizing the reliability of radar by improving the angular resolution. And a lightweight algorithm with a small memory footprint is desired considering that limited computational resources are accessible for automotive radar. Conventionally, iterative algorithms such as iterative shrinkage thresholding algorithm (ISTA) were used for DOA estimation. However, algorithms like ISTA can require many iterations to converge, and a lot of manual parameter tuning is required to obtain optimal performance. Learned ISTA (LISTA) has been used to approximate ISTA with fewer iterations without the necessity of manual tuning by unfolding the iterative algorithm as a neural network. But directly using LISTA is not suitable for DOA estimation due to the large size of the matrices that need to be learned. The large number of learning parameters require a lot of training data, a long training time, and heavy computation. This work proposes to use circular convolutions to reduce the number of learning parameters in the model as well as computation. We show that the circular convolution-based ISTA has better performance metrics than the traditional ISTA.