A 6-DOFs event-based camera relocalization system by CNN-LSTM and image denoising. (15th May 2021)

• Record Type: Journal Article
• Authors: Jin, Yifan; Yu, Lei; Li, Guangqiang; Fei, Shumin
• Is Part Of: Expert systems with applications. Volume 170(2021)
• Publication Date: 2021-05-15
• Subjects: Event image -- Camera relocalization -- Image denoising -- 6-DOFs pose -- Convolutional neural network; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2020.114535
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: The combination of CNN and LSTM camera relocalization method is proposed. The model can predict the 6-DOFs pose with smaller errors at a faster speed. It can enhance the robustness and improve real-time performance of the model. Abstract: At present, in the research of simultaneous localization and mapping systems, many traditional relocalization methods have been replaced by camera relocalization techniques based on convolutional neural network (CNN) and long and short-term memory (LSTM). However, in a system using an event dataset to train the neural network, the complex scenes are chaotic, and the noise of the event images is excessive. Both issues make the model unable to return to the six-degrees of freedom (6-DOFs) pose well. This paper proposes a 6-DOFs pose camera relocalization method based on the CNN image denoising model and CNN-LSTM. Firstly, the CNN image denoising model is used to solve the problem of excessive noise points in complex scenes. Then, a network framework combining CNN and LSTM trains the event camera relocalization model to obtain better 6-DOFs pose accuracy. Finally, the study performs experimental simulations by using complex scene datasets without and with denoising images. Experimental results show that the proposed method of camera relocalization has many advantages. It enhances the robustness of the model in the training process, reduces the mutation situation, and the trained model has a smaller error and faster speed when predicting the pose, thus improving the accuracy and real-time of the camera relocalization model.