Global Localization Using Object Detection in Indoor Environment Based on Semantic Map

Global Localization in indoor environment remains a challenging problem since the indoor environment is always sparse and similar. At the startup stage of the robot, the particle filter algorithm, a widely used localization method, may converge particles to several different positions because of the similar structures in indoor environment, resulting in localization failures. In this paper, a global localization method combining maximum likelihood estimation (MLE) and single shot detector (SSD) detection network is proposed. Firstly, we build an indoor dataset consisting of 10 classes of objects and train a SSD network, which has the best tradeoff between speed and accuracy in object detection field. Based on the well trained model, a semantic map combining 2D grid map generated by Gmapping (a variant of the SLAM algorithm) and object position calculated by SSD network is built. Finally, under the consideration of computation, MLE algorithm and space pyramid method are applied to the process of global localization with object detection. The proposed method is verified with robustness in the experiments and could be easily applied in other localization systems.