DART: Implicit Doppler Tomography for Radar Novel View Synthesis

Tianshu Huang,John Miller,Akarsh Prabhakara,Tao Jin,Tarana Laroia,Zico Kolter,Anthony Rowe

2024-03-07

Abstract:Simulation is an invaluable tool for radio-frequency system designers that enables rapid prototyping of various algorithms for imaging, target detection, classification, and tracking. However, simulating realistic radar scans is a challenging task that requires an accurate model of the scene, radio frequency material properties, and a corresponding radar synthesis function. Rather than specifying these models explicitly, we propose DART - Doppler Aided Radar Tomography, a Neural Radiance Field-inspired method which uses radar-specific physics to create a reflectance and transmittance-based rendering pipeline for range-Doppler images. We then evaluate DART by constructing a custom data collection platform and collecting a novel radar dataset together with accurate position and instantaneous velocity measurements from lidar-based localization. In comparison to state-of-the-art baselines, DART synthesizes superior radar range-Doppler images from novel views across all datasets and additionally can be used to generate high quality tomographic images.

Computer Vision and Pattern Recognition,Machine Learning

What problem does this paper attempt to address?

The problem addressed in this paper is how to achieve new radar view synthesis to create realistic radar range Doppler images and overcoming the challenges faced by existing radar simulation tools in terms of precise scene, material properties, and radar synthesis function models.

DART (Doppler-Assisted Radar Tomography) is an approach inspired by the neural radiation field, which utilizes the unique physical principles of radar to construct a rendering pipeline based on reflectivity and transmissivity for range Doppler imaging. The paper evaluates the performance of DART through a customized data collection platform and radar dataset, as well as precise positional and instantaneous velocity measurements of the radar. Compared to the existing state-of-the-art baseline, DART is able to generate higher-quality radar range Doppler images from a new perspective across all datasets and can generate high-quality tomographic images.

DART addresses the problem of applying neural networks to millimeter-wave radar imaging and overcomes the unique characteristics of radar wave propagation, such as radial propagation and poorer spatial resolution. It utilizes range Doppler representation to reduce angular ambiguity and improves angular resolution by considering relative velocity. DART is able to accurately synthesize radar images from a new perspective by implicitly learning the reflectivity and transmissivity of the scene, rather than just memorizing input data.

The main contributions of the paper are:
1. Proposing a method for radar new view synthesis using range Doppler imaging.
2. Formulating the forward rendering of radar range Doppler images explicitly based on neural radiation field technology and implicitly inverting it to learn neural implicit representation.
3. Building a data collection platform and radar dataset, as well as precise positional and instantaneous velocity measurements.
4. Demonstrating the superior performance of DART in various scenarios compared to existing techniques.

However, due to the reliance on Doppler, DART is limited to static scenes and requires accurate velocity estimation and a moving radar in its current state.