Initial alignment technology is critical for inertial navigation systems (INS). A key metric for evaluating the superiority of initial alignment methods is the reduction of alignment time while maintaining accuracy. Dual-antenna aided INS alignment offers a significant advantage by effectively shortening this initial alignment period. However, the heading observation information provided by the dual-antenna subsystem is highly environment-dependent. Mismatch between the assumed observation error noise covariance matrix and its actual value can degrade alignment accuracy or even lead to alignment failure. To address this challenge, this paper proposes a novel dual-antenna aided inertial initial alignment method. Leveraging observation feedback theory, the method achieves online variance estimation of dual-antenna observation noise. This capability effectively suppresses the adverse impact of time-varying observation noise on both the accuracy and reliability of the initial alignment process. Simulation analysis demonstrates that the proposed method successfully mitigates the influence of observation noise variations and solves the problem of online identification for the statistical characteristics of dual-antenna heading angle errors. Furthermore, while accomplishing the initial attitude alignment of the INS, the method simultaneously provides effective estimates of specific inertial sensor errors within the alignment timeframe.