Application and Analysis of Recurrent Convolutional Neural Network in Visual Odometry

Abstract

Visual odometry is an important part of simultaneous localization and mapping. Now it is widely used in driverless, robots and other fields. The traditional visual odometry is composed of many complex modules, such as feature extraction, camera calibration and other modules, only great cooperation between various modules can achieve good results. In addition, the high complexity, high hardware requirements and weak anti-interference ability make it difficult for the visual odometry to achieve the expected effect. We present a visual odometry by using Recurrent Convolutional Neural Network in this paper, which inputs image sequences, obtains pose data then outputs. Its complexity is lower than that of traditional algorithms. Results show that this algorithm can estimate the camera odometry end-to-end, has higher accuracy, stability and lower complexity.