A deep learning framework for robust semantic SLAM

Semantic simultaneous localization and mapping (SLAM) is susceptible to several sources of noise that hinder the accuracy of its trajectory and map estimates. Such sources include inaccurate landmark pose estimation and sensor limitations. In this paper, a novel deep learning based approach is proposed to improve the accuracy of semantic SLAM by reducing the trajectory estimation error. A deep neural network consisting of various non-linear activation functions is structured and pre-trained by means of an unsupervised greedy layer-wise pre-training technique. The network is then fine-tuned using the adaptive moment estimation (Adam) optimizer. The training datasets were collected using several simulated and realtime experiments and are composed of two parts, the estimated trajectory and the corresponding ground truth. Ground truth trajectories were obtained using a motion capture system in realtime experiments. The effectiveness of the proposed approach was shown through simulated experiments, real-time experiments, and a sequence from the Technical University of Munich (TUM) RGB-D dataset. The performance of the deep neural network (DNN) was tested with different pre-training techniques and the proposed unsupervised greedy layer-wise pre-training technique proved to perform the best across training, validation, and testing datasets in terms of reducing the mean absolute trajectory error (ATE).