Short-Range Target Tracking Using High-Resolution Automotive Radars

Abstract

There is growing interest in the application of high-resolution radars in autonomous vehicles due to their affordability and high angular resolution. However, the azimuth ambiguity caused by the large physical distance between radar antennas relative to the signal wavelength is a challenge for its application. The problem of multiple extended target tracking using high-resolution radar measurements with azimuth ambiguity is considered. A novel pseudo-3D assignment (P3DA) method based on the pseudo measurement set (PMS) is proposed to resolve the azimuth ambiguity. This method can resolve mono (single) and split (duplicated) azimuth ambiguities common in extended target tracking. The Lagrangian relaxation based on a flexible search (LR-FS) algorithm is proposed to solve the P3DA-PMS problem efficiently. Simulation and experiment results show that the proposed algorithm outperforms conventional methods that do not address the azimuth ambiguity of extended target tracking. Since data association with only one data frame will lose information about target evolution and cannot change an association later based on subsequent measurements, a novel two-step multiframe assignment method is proposed to resolve split and azimuth ambiguity separately. In the first step, the split ambiguity is resolved by the PMS-to-PMS association, resulting in a merged PMS (MPMS). In the second step, the azimuth ambiguity is resolved by the Track-to-MPMS association. Numerical results show that the proposed method performs better than the P3DA-PMS-based method. The vehicles tracking with high-resolution radars need to provide information about their orientation and shape to achieve lidar-like performance. Due to self-occlusion, the L-shape model is frequently utilized to depict the structure of a typical vehicle. Since the measurement accuracy of high-resolution radars is not as high as that of lidars, radar measurement noise cannot be ignored. Moreover, as a side effect of using large wavelengths, multiple measurements may be produced per time step due to multipath effects. As a result, more outliers and inliers can be generated in high-resolution radar measurements. A novel lognormal likelihood-aided L-shape model is proposed to approximate the distribution of high-resolution radar measurements of vehicles. Numerical results evaluated on simulation data and the KITTI dataset show that the proposed algorithm achieves smaller orientation and position errors and larger generalized intersection over union (GIoU) compared to existing L-shape fitting algorithms for lidar measurements.