Global Navigation Satellite System (GNSS) and Inertial Navigation System (INS) are two most widely used navigation systems. They both have their strengths and weaknesses, which naturally leads to the idea of combining them into single solution. The main goal of the research was to fuse data from multiple sensors to improve the robustness of the pedestrian navigation in urban environment. To achieve the goal, we focus on the data preprocessing and its effect on the result, solve the velocity drift problem when integrating raw acceleration and implement Kalman filter with velocity correction using Long Short-Term Memory model. The LSTM model, trained on accelerometer and GNSS data, achieved mean average error of 0.087 m/s and R2 value of 0.85 in the task of speed prediction on test dataset. To mitigate velocity drift inherent in inertial navigation, the LSTM-predicted speed is combined with the Kalman filter’s velocity estimate, compensating for sensor noise and bias. It was demonstrated that we can restore original trajectory when GNSS signal is jammed or unavailable for up to 10 s. However, the accuracy of the fused trajectory remains constrained by unreliable orientation estimate from magnetometer sensor.