To address the challenge of accurately determining the environmental region of unmanned vehicles during seamless indoor-outdoor positioning, a regional recognition method for seamless indoor-outdoor localization is proposed. Firstly, a joint prediction model integrating particle swarm optimization-support vector machine (PSO-SVM) and hidden Markov model (HMM) is designed. Environmental feature data acquired by sensors serve as model inputs to generate regional recognition results. Secondly, three environmental models are introduced to describe the vehicle's operational environment, with corresponding measurement information selected based on the regional recognition outcomes. Finally, the regional transition probabilities are utilized to update the switching probabilities of the three environmental sub-models in the interactive multiple model (IMM) algorithm, thereby enhancing the accuracy of environmental region recognition and positioning precision for seamless indoor-outdoor navigation. The results of real-vehicle experiment show that the proposed joint recognition method achieves an accuracy of 98.09% in region recognition, representing improvements of 2.13% and 9.53% compared to using PSO-SVM or HMM alone. Further experiments indicate that the proposed seamless positioning method enhances the average positioning accuracy by 43.75% and 22.30% compared to the traditional federated Kalman filter (FKF) algorithm and IMM algorithm, respectively.

Traditional neural network algorithms are prone to consuming a long time and getting stuck in local optima when artificial intelligence methods are applied to geomagnetic navigation and positioning. To address these issues, a method for geomagnetic positioning of aircraft based on improved gradient-based genetic algorithm optimized extreme learning machine neural network (GGA-ELM) is proposed. The training efficiency is greatly improved based on the optimized ELM network and the risk of falling into local optimum is effectively reduced as well by introducing an elite reverse learning strategy into the traditional genetic algorithm. Some aeromagnetic data measured by drone are used for investigation. The experimental results show that the training time of the GGA-ELM model is significantly reduced compared with the CNN, BiLSTM and LSTM models. In addition, the localization error of the GGA-ELM model is about 4 m, and the localization time is 0.003 s. Compared with the ELM, GA-ELM, CNN, BiLSTM, RBF and LSTM models, based on the GGA-ELM method, the localization accuracy is improved by 86.6%, 115.9%, 417.8%, 187.6%, 216.5%, and 107.5%, respectively. The localization time is reduced up to 0.947 s. From the results, it is clearly seen that the proposed method has better positioning stability and higher accuracy on aircraft localization.