Visual Robot Relocalization Based on Multi-Task CNN and Image-Similarity Strategy

The traditional CNN for 6D robot relocalization which outputs pose estimations does not interpret whether the model is making sensible predictions or just guessing at random. We found that convnet representations trained on classification problems generalize well to other tasks. Thus, we propose a multi-task CNN for robot relocalization, which can simultaneously perform pose regression and scene recognition. Scene recognition determines whether the input image belongs to the current scene in which the robot is located, not only reducing the error of relocalization but also making us understand with what confidence we can trust the prediction. Meanwhile, we found that when there is a large visual difference between testing images and training images, the pose precision becomes low. Based on this, we present the dual-level image-similarity strategy (DLISS), which consists of two levels: initial level and iteration-level. The initial level performs feature vector clustering in the training set and feature vector acquisition in testing images. The iteration level, namely, the PSO-based image-block selection algorithm, can select the testing images which are the most similar to training images based on the initial level, enabling us to gain higher pose accuracy in testing set. Our method considers both the accuracy and the robustness of relocalization, and it can operate indoors and outdoors in real time, taking at most 27 ms per frame to compute. Finally, we used the Microsoft 7Scenes dataset and the Cambridge Landmarks dataset to evaluate our method. It can obtain approximately 0.33 m and 7.51 [Formula: see text] accuracy on 7Scenes dataset, and get approximately 1.44 m and 4.83 [Formula: see text] accuracy on the Cambridge Landmarks dataset. Compared with PoseNet, our CNN reduced the average positional error by 25% and the average angular error by 27.79% on 7Scenes dataset, and reduced the average positional error by 40% and the average angular error by 28.55% on the Cambridge Landmarks dataset. We show that our multi-task CNN can localize from high-level features and is robust to images which are not in the current scene. Furthermore, we show that our multi-task CNN gets higher accuracy of relocalization by using testing images obtained by DLISS.