In automotive radar systems, the range and Doppler velocity of vehicles surrounding the radars can be estimated by performing a fast Fourier transform (FFT) on the processed received signals reflected by the vehicles. The trade-off between unambiguity for estimation and resolution in FFT-based estimation methods can be broken with low computational complexity by introducing the Chinese remainder theorem (CRT). However, there are two challenges in CRT-based methods: the additional target association procedure and the error propagation drawback. In this study, a novel multi-waveform radar frame structure is proposed to facilitate the use of the CRT. Based on the frame structure, a corresponding CRT-based target association method is proposed to eliminate ghost targets. Moreover, a generative adversarial neural network (GAN)-based target association method is proposed to further address the error propagation drawback. Simulation results show the robustness of the GAN-based method, with an outstanding performance compared to other rule-based methods, even in severe error scenarios.