Vehicle Localization Method Based on Multitask Cascaded Learning During GPS Outages

Enhancing the efficiency of the inertial navigation system (INS)/global positioning system (GPS) integrated navigation systems for land vehicles during GPS outages is exceedingly challenging. Existing research commonly model vehicle motion as a single scenario. However, vehicle motion encompasses diverse behaviors such as stopping, turning, and moving straight, each presenting distinct dynamic features and environmental conditions. To improve the positioning accuracy of vehicles in complex urban environments, we propose a multitask sequential cascaded learning (CL) model, which primarily comprises three components: 1) a vehicle stationary detector that utilizes a deep residual convolutional neural network (DRCNN) to accurately detect the stationary status of vehicles, promptly correcting positioning errors when the vehicle is parked; 2) a turning detector that utilizes a threshold-based approach to accurately detect vehicle turning states, effectively categorizing vehicle motion into turning and moving straight scenarios, thus simplifying the complexity of the vehicle positioning system; and 3) a predictor that utilizes a dual temporal-spatial adaptive neural network (D-TSANN) to accurately predict pseudo-GPS position, effectively bridging the GPS gaps. The D-TSANN consists of two temporal-spatial adaptive neural networks (TSANNs) with similar structures but distinct learning strategies, specifically addressing turning and moving straight scenarios. These TSANNs incorporate CNN, gated recurrent unit (GRU) neural network, and self-attention (SA) modules to autonomously extract spatiotemporal features from the original inputs. Road test results demonstrate that during GPS outages, the proposed CL model outperforms existing deep learning algorithms, providing a more stable and reliable solution for vehicle navigation.