Occupancy grid mapping is a common approach to support automotive driving perception capabilities. We present an occupancy grid estimation algorithm using sensor point-cloud measurements aided by side information from other sensing modalities like cameras. This prior side information is in the form of an erroneous occupancy map estimate, referred to as prior support information. Specifically, we extract a prior map using you only look once (YOLO) object detection on camera images. A sparse Bayesian learning-based mapping algorithm is designed with a modified hierarchical model to incorporate this prior. Experiments done on public real-world driving datasets, nuScenes and RADIATE, demonstrate that our approach achieves better target detection and scatter noise reduction than the state-of-the-art methods. Furthermore, our method seamlessly works on the two datasets although we train YOLO only using camera images from nuScenes.

We consider the problem of generating automotive radar super-resolution maps from low-resolution radar maps and camera images. This problem is relevant in automotive driving for synthetic sensor data generation to support improved environmental perception. We propose a radar super-resolution sensing approach based on multimodal data fusion between low-resolution radar range-azimuth (RA) maps and aligned camera images. Our method employs a U-Net-based autoencoder architecture enhanced with visual features extracted from a pre-trained ResNet50 encoder, enabling the model to generate high-resolution RA maps that approximate ground truth radar data. We evaluate the proposed method on the RADIal and RaDICaL datasets, which cover diverse driving environments and radar configurations. Quantitative and qualitative results demonstrate that our approach outperforms a baseline model and prior state-of-the-art methods, particularly in resolving fine spatial details in scenarios with closely-spaced vehicles and pedestrians.

Digital radars with low-resolution analog-to-digital converters (ADCs) have attracted attention as a solution to reducing the high digital processing complexity and power consumption at the receiver. Radars employing low-resolution ADCs, however, have a limited dynamic range, due to which high-radar cross section (RCS) targets mask low-RCS targets. The masking occurs because the quantized output is primarily determined by returns from high-RCS targets. To enhance the dynamic range of such radars, we propose to operate the ADC at a high resolution in the initial slow-time slot of each radar frame. The resulting high-resolution measurements are used together with the known Doppler statistics of dominant targets to construct a dither signal, which is used as a quantization threshold to acquire low-resolution ADC measurements in the subsequent slow-time slots. By incorporating situation awareness in the form of Doppler statistics, our dither signal can suppress returns from strong targets, effectively unmasking weak targets with low-resolution measurements. We analyze system performance in terms of the probability of detection and show that the proposed approach outperforms existing methods in enhancing the detection of weak targets. The simulations demonstrate that our method significantly improves target detection and reduces the normalized mean square error (NMSE) in the estimated radar channel over comparable benchmarks.

Light detection and ranging (LiDAR) is used in robots and in automotives to obtain the perception of the surrounding environment. Traditional spinning LiDARs scan the environment uniformly along all angular directions by operating at a constant rotational speed, with fixed sensing parameters throughout a rotation. Such a sensing approach, however, is suboptimal when information about static obstacles in the environment is available at the LiDAR. In this work, we introduce ELLAS, a first-of-its-kind spinning LiDAR system that dynamically adapts its range and resolution over the field of view. This adaptation is achieved by optimizing the ranging parameters at the LiDAR and the instantaneous rotational speed of the spinning platform to the location of static objects in scene topology maps. With the optimized settings, ELLAS results in a longer range along directions where static obstacles are farther away and achieves a higher angular resolution around those directions.

Occupancy maps are used in automotive driving applications to understand the scene around the vehicle using sensor data like LiDAR measurements. State-of-the-art work relies on pattern-coupled sparse Bayesian learning (PCSBL) to estimate the occupancy map from LiDAR point cloud data by leveraging spatial dependencies across grids in the map. However, PCSBL has high computational complexity, posing challenges for real-time implementation on large-sized grid maps. In this work, we propose two methods to improve the computational efficiency of PCSBL for occupancy grid mapping by exploiting the narrow angular interactions of sensor measurements with the map. The first method partitions the measurements into spatially disjoint submaps that can be processed in parallel. The second method exploits the angular structure to impose a block structure on the measurement matrix, allowing more efficient matrix computations to accelerate the algorithm. Experiments on the nuScenes public dataset show that the presented methods reduce computational runtime compared to the benchmark PCSBL method while preserving detection accuracy.

An occupancy grid map (OGM) is used in automotive driving applications to model the vehicle surroundings using data from sensors on vehicles like light detection and ranging (LiDAR), radar, or their fusion. In stateof- the-art work, pattern-coupled sparse Bayesian learning (PCSBL) was used to estimate the OGM by leveraging spatial dependencies in the map when either a single sensor modality was used or when fusion of multiple sensor modalities was employed. The PCSBL method, however, has high computational complexity, making real-time implementation challenging for large-sized grid maps. To address this limitation, we propose several methods to improve the computational efficiency of PCSBL while maintaining mapping accuracy. First, we utilize a precomputed lookup table to accelerate selection matrix construction. Second, we implement adaptive resolution reduction based on sensor measurements, lowering problem dimensionality where coarse resolution suffices. Third, we develop two novel methods that exploit the narrow angular interactions between measurements and the map regions to enhance computational efficiency. The first method partitions measurements into spatially disjoint submaps that enable parallel processing. The second method exploits the angular structure to impose a block structure on the selection matrix, reducing the computational overhead of matrix operations. Experiments on the nuScenes and RADIATE public datasets show that the presented methods reduce runtime by at least an order of magnitude compared with the benchmark PCSBL and fusion-based PCSBL methods while preserving detection accuracy.

Phase-modulated continuous-wave (PMCW) radars offer better multiple antenna support and flexible waveform design. These radars, however, suffer from the high computational complexity of correlation processing used to obtain the range of targets, particularly due to the use of long code sequences. To address this, we exploit signal sparsity and propose an adaptive sliding block-based fast Fourier transform (FFT) correlator that selectively processes only the relevant range bins and dynamically aligns processing blocks using a time-shift parameter. This sliding mechanism minimizes boundary effects and reduces the number of blocks required for processing. The framework also incorporates external sensor data for context-aware range-bin selection and includes an analytical formulation for optimal block size selection. Simulations demonstrate that the proposed method preserves the signal-to-noise ratio (SNR) and target peaks while significantly reducing processing time. The effectiveness of the proposed method is demonstrated through numerical simulations.

Radar is a key technology in automotive driving for target detection and perception. In this work, we leverage prior environmental information in the form of occupancy maps to design space-time codes for a fully digital MIMO radar. We transform this design problem into the optimization of spatial beamforming gains and time-domain codes. The beamforming gains are optimized to enhance the strength of returns from cells associated with a higher uncertainty of occupancy. The timedomain codes are optimized to minimize the correlation between returns of targets within the drivable space. We validate our method on the nuScenes dataset to show that the designed spacetime codes achieve higher detection rates than designs that do not rely on prior information from occupancy maps.

We tackle the problem of estimating a binary occupancy grid map by fusing point cloud data from LiDAR and radar sensors for automotive driving perception. To this end, we introduce two sparsity measurement models for fusion, formulating occupancy mapping as a sparse binary vector reconstruction problem. The first model jointly estimates a common map from all measurements, while the second assumes a shared map and an innovation component for each modality's measurements. We use the pattern-coupled sparse Bayesian learning algorithm to recover maps, leveraging the inherent sparsity and spatial dependencies in automotive occupancy maps. Numerical experiments on the RADIATE public dataset show that our fusion-based approach improves mapping accuracy compared to single-modality and high-level fusion mapping algorithms.

Millimeter-wave radar is a common sensor modality used in automotive driving for target detection and perception. These radars can benefit from side information on the environment being sensed, such as lane topologies or data from other sensors. Existing radars do not leverage this information to adapt waveforms or perform prior-aware inference. In this paper, we model the side information as an occupancy map and design transmit beamformers that are customized to the map. Our method maximizes the probability of detection in regions with a higher uncertainty on the presence of a target. Simulation results on the nuScenes dataset show that the designed beamformer achieves substantially higher detection rates than a conventional omnidirectional beamformer for the same transmitted power.

Automotive perception involves understanding the external driving environment and the internal state of the vehicle cabin and occupants using sensor data. It is critical to achieving high levels of safety and autonomy in driving. This article provides an overview of different sensor modalities, such as cameras, radars, and light detection and ranging (LiDAR) used commonly for perception, along with the associated data processing techniques. Critical aspects of perception are considered, such as architectures for processing data from single or multiple sensor modalities, sensor data processing algorithms and the role of machine learning techniques, methodologies for validating the performance of perception systems, and safety. The technical challenges for each aspect are analyzed, emphasizing machine learning approaches, given their potential impact on improving perception. Finally, future research opportunities in automotive perception for their wider deployment are outlined.