In recent years, integrated polarized light/inertial heading measurement systems have been widely used to obtain autonomous heading measurements of small unmanned combat platforms in the case of satellite navigation rejection. However, existing polarized light/inertial heading measurement systems have certain limitations. For example, they can only measure the heading angle in environments where continuous observations can be obtained. When encountering a complex environment with trees and/or tall buildings, the measured heading angle will contain sharp noise which greatly affects its accuracy. In particular, when encountering an underpass, it will lead to the complete loss of lock of the polarized light compass signal. Therefore, for the problem of sharp noise arising from a complex environment, a robust Cubature Kalman filter (CKF) data-fusion algorithm is proposed and verified by experiments. It is proved that the robust CKF algorithm has a certain ability to filter out the effects of poor measurements. After application of the robust CKF algorithm, the Root Mean Square Error (RMSE) of the heading angle reaches 0.3612°. This method solves the problem of low precision and poor stability of the polarized light/inertial system when high buildings and/or trees are contained in a complex environment. Next, in view of the problem that the polarized light compass signal is completely lost due to passing through an underground passage, a random forest regression (RFR) neural network model is established and introduced into the combined system. Simulated and outdoor experiments are carried out to verify the designed model using data obtained with a vehicle. The RMSE of the heading angle obtained in the experiment is 1.1894°, which solves the problem that the polarized light/inertial system cannot utilize discontinuous observations when attempting to detect the carrier heading angle.