We present a novel mmWave radar-based solution for robust, non-contact estimation of heart rate (HR) and respiratory rate (RR) under dynamic conditions. Our approach begins with a targeted denoising pipeline—comprising background clutter subtraction, Kalman filtering, and Variational Mode Decomposition (VMD)—to mitigate environmental interference, hardware-induced noise, and low-frequency perturbations arising from subject motion. We then introduce CLANet (Convolution-LSTM-Attention Network), a hybrid neural architecture that fuses local feature extraction, temporal dependency modeling, and attention-based emphasis on salient signal segments. Extensive experiments on a 14-participant dataset, which includes diverse motion and conversational tasks, demonstrate that CLANet outperforms state-of-the-art baselines in both accuracy and robustness, consistently providing high-fidelity HR/RR detection. Our framework lays the groundwork for reliable, real-time vital sign monitoring with mmWave radar, offering a promising alternative to contact-based sensors across healthcare, human–computer interaction, and other human-centric domains.