In a complex urban environment, when the GNSS signal fails, the existing methods for estimating the vehicle pose using a monocular camera or an inertial navigation system suffer serious cumulative errors. To address these issues, a vehicle pose fusion estimation algorithm based on the cascaded deep neural network ( CDNN) is proposed. First, CDNN is designed to reduce the cumulative error caused by scale blur and scale drift in monocular cameras. Secondly, to reduce the introduced device noise, a simplified inertial sensor system (RISS) is used to obtain the vehicle lateral and longitudinal acceleration and yaw rate. To reduce the influence of uncertain noise in the system, H∞ filtering is used to fuse the outputs of CDNN and RISS to accurately estimate the vehicle pose while keeping high-frequency output. Compared with the method based on the Kalman filter, experimental results on the KITTI dataset show that, the root mean square error (RMSE) of the easting position estimated by the proposed algorithm is reduced by 41. 3% , and the RMSE of the estimated northing position is reduced by 70. 6% , and the RMSE of the estimated heading angle is reduced by 6 K 6 e . y 6 w % o .