Radar-based object detection is essential for autonomous driving due to radar's long detection range. However, the sparsity of radar point clouds, especially at long range, poses challenges for accurate detection. Existing methods increase point density through temporal aggregation with ego-motion compensation, but this approach introduces scatter from dynamic objects, degrading detection performance. We propose DoppDrive, a novel Doppler-Driven temporal aggregation method that enhances radar point cloud density while minimizing scatter. Points from previous frames are shifted radially according to their dynamic Doppler component to eliminate radial scatter, with each point assigned a unique aggregation duration based on its Doppler and angle to minimize tangential scatter. DoppDrive is a point cloud density enhancement step applied before detection, compatible with any detector, and we demonstrate that it significantly improves object detection performance across various detectors and datasets.

Point clouds are unordered sets of coordinates in 3D with no functional relation imposed on them. The Rigid Transformation Universal Manifold Embedding (RTUME) is a mapping of volumetric or surface measurements on a 3D object to matrices, such that when two observations on the same object are related by a rigid transformation, this relation is preserved between their corresponding RTUME matrices, thus providing linear and robust solution to the registration and detection problems. To make the RTUME framework of 3D object detection and registration applicable for processing point cloud observations, there is a need to define a function that assigns each point in the cloud with a value (feature vector), invariant to the action of the transformation group. Since existing feature extraction functions do not achieve the desired level of invariance to rigid transformations, to the variability of sampling patterns, and to model mismatches, we present a novel approach for designing dense feature extraction functions, compatible with the requirements of the RTUME framework. One possible implementation of the approach is to adapt existing feature extracting functions, whether learned or analytic, designed for the estimation of point correspondences, to the RTUME framework. The novel feature-extracting function design employs integration over $SO(3)$ to marginalize the pose dependency of extracted features, followed by projecting features between point clouds using nearest neighbor projection to overcome other sources of model mismatch. In addition, the non-linear functions that define the RTUME mapping are optimized using an MLP model, trained to minimize the RTUME registration errors. The overall RTUME registration performance is evaluated using standard registration benchmarks, and is shown to outperform existing SOTA methods.

Object detection in radar imagery with neural networks shows great potential for improving autonomous driving. However, obtaining annotated datasets from real radar images, crucial for training these networks, is challenging, especially in scenarios with long-range detection and ad-verse weather and lighting conditions where radar performance excels. To address this challenge, we present Rad-SimReal, an innovative physical radar simulation capable of generating synthetic radar images with accompanying annotations for various radar types and environmental conditions, all without the need for real data collection. Re-markably, our findings demonstrate that training object de-tection models on RadSimReal data and subsequently eval-uating them on real-world data produce performance lev-els comparable to models trained and tested on real data from the same dataset, and even achieves better performance when testing across different real datasets. Rad-SimReal offers advantages over other physical radar simulations that it does not necessitate knowledge of the radar design details, which are often not disclosed by radar sup-pliers, and has faster run-time. This innovative tool has the potential to advance the development of computer vision al-gorithms for radar-based autonomous driving applications. Our GitHub: https://yuvalhg.github.io/RadSimReal.

We consider the problems of estimating the underlying transformation and the detection of 3-D objects undergoing rigid transformations. It has been shown that the Rigid Transformation Universal Manifold Embedding (RTUME) provides a mapping from the set of all possible observations on some object to a transformation covariant matrix representation, such that its column space is invariant to the geometric transformation. In this paper, we re-derive and adapt the RTUME for the case where the observations are in the form of meshed surfaces. We prove that by evaluating the integrals that define the RTUME operator as surface integrals on the mesh representation of the observed surface, the invariance and covariance properties of the RTUME matrix representation hold, similarly to the case of point cloud observations. It is shown that the RTUME matrix representation can be efficiently evaluated using a barycentric coordinate representation of the observed surface mesh representation. The proposed Mesh-RTUME is shown to outperform the RTUME representation, evaluated from the point cloud representation of the surface, both in transformation estimation and in keypoint detection.

Automotive radars have an important role in autonomous driving systems. The main challenge in automotive radar detection is the radar’s wide point spread function (PSF) in the angular domain that causes blurriness and clutter in the radar image. Numerous studies suggest employing an ’end-to-end’ learning strategy using a Deep Neural Network (DNN) to directly detect objects from radar images. This approach implicitly addresses the PSF’s impact on objects of interest. In this paper, we propose an alternative approach, which we term ’’Boosting Radar Reflections" (BoostRad). In BoostRad, a first DNN is trained to narrow the PSF for all the reflection points in the scene. The output of the first DNN is a boosted reflection image with higher resolution and reduced clutter, resulting in a sharper and cleaner image. Subsequently, a second DNN is employed to detect objects within the boosted reflection image. We develop a novel method for training the boosting DNN that incorporates domain knowledge of radar’s PSF characteristics. BoostRad’s performance is evaluated using the RADDet and CARRADA datasets, revealing its superiority over reference methods.

We consider the problem of classifying 3-D objects undergoing rigid transformations. It has been shown that the rigid transformation universal manifold embedding (RTUME) provides a mapping from the orbit of observations on some object to a single low-dimensional linear subspace of Euclidean space. This linear subspace is invariant to the geometric transformations. In the classification problem the RTUME subspace extracted from an experimental observation is tested against a set of subspaces representing the different object manifolds, in search for the nearest class. We elaborate on the design problem of the RTUME operator in the case where the point cloud sampled from the object is sparse, noisy, and non-uniformly sampled. By introducing metric learning and negative-mining techniques into the framework of Grassmannian dimensionality reduction for universal manifold embedding, we improve classification performance for these challenging sampling conditions.