Abstract: Under GNSS-denied environments, the inertial navigation of smartphones faces significant cumulative errors. Traditional non-holonomic constraints (NHC) are also ineffective during dynamic maneuvers such as turns. To address these issues, this paper proposes a deep learning-assisted NHC enhancement method. This approach utilizes a fusion model of a one-dimensional convolutional residual network (ResNet-1D) and a long short-term memory (LSTM) network to accurately recognize vehicle straight and turning states from smartphone inertial measurement unit (IMU) data. This recognition enables dynamic adjustment of the observation noise covariance in the NHC of the extended Kalman filter (EKF), strengthening constraints during straight driving and relaxing them during turns. Experimental results show that the proposed state recognition model improves turning recognition accuracy from about 65% to approximately 95%. Furthermore, real-world test data demonstrate that, under GNSS-denied conditions, the method achieves a 36.4% to 74.6% improvement in horizontal relative positioning accuracy compared to the traditional Z-axis gyroscope integration method. This study enhances smartphone navigation performance in complex environments through intelligent motion state recognition and adaptive NHC adjustment, offering a new approach for low-cost, high-precision positioning.