With the development of indoor positioning technology, the Inertial Measurement Unit (IMU) in smartphones has been widely applied to vehicle speed estimation. To improve the accuracy of vehicle speed estimation in complex underground parking environments and the user experience of indoor navigation with the in-vehicle smartphone, this paper proposes a fusion model combining a 1D Convolutional Neural Network (1DCNN) and Long Short-Term Memory (LSTM) network based on smartphone IMU data. The model combines a 1DCNN for spatiotemporal feature extraction and uses LSTM to handle temporal dependencies, significantly enhancing the accuracy of speed estimation. Experimental results show that, in four different underground parking scenarios, the model’s average error is 0.741 m/s, significantly outperforming traditional methods. The study demonstrates the broad application potential of deep learning models based on smartphone IMUs, particularly in intelligent transportation and autonomous driving fields.