SqueezeNet-Based Range, Angle, and Doppler Estimation for Automotive MIMO Radar Systems

The frequency modulated continuous waveform multiple - input multiple - output (FMCW MIMO) radar is of great interest to the automotive industry that provides high-end automobiles equipped with parking assistance, lane departure warning, and adaptive cruise control. These radars can simultaneously detect the range, angle, and doppler of the surrounding objects, such as cars, trucks, bicycles, and pedestrians and relay this information to the central control to provide a safe and collision-free cruise control for the self-driving vehicle. The traditional approach for the radar range, angle, and Doppler estimations is the Fast Fourier Transform (FFT)FFT which is computationally efficient but suffers from poor angular resolution. On the other hand, high-resolution techniques such as the Multiple SIgnal Classifier (MUSIC), the Estimation of Signal Parameters via Rotational Invariance Techniques (ESPRIT), and the Minimum Variance Distortionless Response (MVDR) can achieve more accurate estimations but are computationally expensive. Moreover, these high-resolution techniques are very sensitive to clutter and interferences and cannot effectively distinguish targets from clutters in such an environment. In this paper, we propose a deep-learning- based FMCW MIMO radar in which the range, angle, and Doppler estimation are treated as a multilabel classification problem. The deep - learning approach is based on the SqueezeNet transfer learning approach to overcome the limitations on the amount of training data and training time. Simulation results demonstrate that the proposed approach outperforms the MVDR method in the presence of clusters and jammers and can achieve a high angular resolution of 2 degrees.