HyDNN-PDR: A Novel Hybrid Deep Neural Network Based Pedestrian Dead Reckoning Navigation System in Challenging Terrains

The modernization of live training requires accurate dismounted soldier trainee position/tracking to enable the convergence of Live, Virtual, and Constructive (LVC) training environments. The position jitter or the random position changes make the training unreliable under the condition of GPS satellite signal attenuation or blockage. In this project, a novel artificial neural network based pedestrian dead reckoning navigation system is proposed to upgrade the real-time navigation information accurately even in a variety of challenging terrains. To provide accurate location information, the designed pedestrian dead reckoning navigation platform utilizes an innovative online and offline hybrid machine learning based sensing data processing technique to be applied to the portable, low stringent size, weight, and power constraints (SWaP-C) inertial sensing platform. The hardware platform is composed of multi-inertial sensors, pre-processing filters, a GPU processor and memory. The deep learning based pedestrian dead reckoning navigation algorithms extract and sensing data features and the pedestrian navigation pattern. Then the navigation error would be further reduced by incorporating the novel neural network structure (TS-LSTM). The self-organizing structure and Generative Adversarial Network (GAN) could also make the solution reliable in the scenarios of complex environments and challenging terrains. The preliminary results validate the proposed solution’s feasibility.