ORBDeepOdometry - A Feature-Based Deep Learning Approach to Monocular Visual Odometry

Abstract

Visual Odometry (VO) is an application of computer vision which helps in autonomous navigation of Self-Driving vehicles, Robots etc. The odometry pose of the vehicle can be estimated using the consecutive frames based on monocular or stereo camera setup mounted on the vehicle. The Simultaneous Localization and Mapping (SLAM) based approaches have shown phenomenal results in predicting the vehicle odometry. Deep Learning based approaches are emerging in the field of computer vision and are outperforming other classic algorithms. But, visual odometry is one of the field where deep learning approaches have not yet been exploited largely. This paper proposes the use of Deep Learning based pipeline approach to solve the problem of visual odometry. The paper boosts the performance of deep learning approach by extracting quite more features before training the network. This paper proposes the use of ORB based feature extractor along with Convolutional Neural Network based dimensionality reduction and stacking multiple deep LSTM's for modelling the sequential data. In this case, the KITTI vision Benchmark dataset is used to model the network. The accuracy of the network is examined by calculating the error between the predicted odometry output and the ground truth odometry output. The results are compared with different Convolutional Neural Network (CNN) architectures proposed for the same task. The average translation error from the proposed system is 11.99% and the average rotational error from the proposed system is 0.0462 degrees per meter.