In this paper, a vehicle positioning scheme using the invariant extended Kalman filter (InEKF) and deep learning is studied. Firstly, by introducing the wheel speedometer measurement model, a deep neural network based on autoencoder was constructed, and the true value of vehicle speed was reconstructed. Then, based on InEKF, a filtering algorithm with SE(3) as the state quantity is derived, and the algorithm is used to fuse multi-source information to estimate the vehicle position. Experimental results show that compared with the existing advanced methods, the proposed vehicle positioning system can significantly improve the positioning accuracy in urban environment.