Abstract: Virtual odometer(VODO) and non-holonomic constraint (NHC) can effectively suppress the inertial navigation system (INS) error dispersion during GNSS outage. In recent years, virtual odometer based on machine learning has been widely used in in-vehicle localization, but there are still many limitations on the prediction accuracy and operational efficiency of the existing models. Meanwhile, the fixed noise NHC is difficult to accurately portray complex motion states such as vehicle sideslip. In this paper, we proposes an adaptive INS/NHC localisation method based on the convolutional neuravl network (CNN)/Transformer virtual odometry reinforcement model. The method adopts multi-layer CNN to extract local features in the IMU raw data, captures global timing dependencies with the help of Transformer coding layer, and realises high-performance forward speed budget by efficient parallel computation through multi-attention mechanism; and constructs an adaptive model of NHC noise by combining the predicted forward speed and the vehicle motion state to realise the dynamic adjustment of constraint parameters. The experimental results show that the prediction accuracy of forward speed of the hybrid network reaches 0.56 m/s, and the prediction accuracy and training efficiency are improved by 13.3% and 8.1% respectively compared with the classical network. In the case of 135 seconds of GNSS outages and continuous vehicle turning, the planar positioning accuracy of this paper’s method is improved by 18.85% compared with the traditional constraint method, which effectively suppresses the system error dispersion.