An Adaptive Integrated Positioning Method for Urban Vehicles Based on Multitask Heterogeneous Deep Learning During GNSS Outages

The global navigation satellite system (GNSS)/inertial navigation system (INS) integrated system is widely used in the field of urban vehicle position and navigation due to its low cost and strong complementarity between subsensors; however, in the urban complex environment (such as tunnels and overpasses) where the GNSS signal is completely blocked, the integrated system degenerates to pure INS, resulting in rapid accumulation of errors and significant degradation of positioning performance. In this contribution, an intelligent integrated navigation method based on multitask heterogeneous deep learning and adaptive Kalman filter (AKF) is proposed. The temporal convolutional network (TCN) in the multitask heterogeneous deep learning module outputs the pseudo-GNSS positioning information, and the gated recurrent unit (GRU) network detects the parking state of vehicles in modern cities and generates pseudo-zero-velocity measurements. These pseudo measurements are used to correct INS errors and improve positioning accuracy. The AKF algorithm is used to suppress the pseudo-GNSS position error that accumulates and expands with time, especially when GNSS is outage for a long time. The results of vehicle-mounted road experiments show that heterogeneous deep learning methods can effectively improve pseudo-GNSS position information prediction accuracy and parking state detection accuracy, and AKF can also effectively suppress pseudo-GNSS position information error. Compared with the existing GNSS/INS integrated positioning methods based on deep learning, the proposed method has higher positioning accuracy in position, velocity, and heading, especially when the GNSS is outage for a long time.