Overfitting reduction of pose estimation for deep learning visual odometry

Abstract

Error or drift is frequently produced in pose estimation based on geometric "feature detection and tracking" monocular visual odometry(VO) when the speed of camera movement exceeds 1.5m/s. While, in most VO methods based on deep learning, weight factors are in the form of fixed values, which are easy to lead to overfitting. A new measurement system, for monocular visual odometry, named Deep Learning Visual Odometry(DLVO), is proposed based on neural network. In this system, Convolutional Neural Network(CNN) is used to extract feature and perform feature matching. Moreover, Recurrent Neural Network(RNN) is used for sequence modeling to estimate camera's 6-dof poses. Instead of fixed weight values of CNN, Bayesian distribution of weight factors are introduced in order to effectively solve the problem of network over-fitting. The 18,726 frame images in KITTI dataset are used for training network. This system can increase the generalization ability of network model in prediction process. Compared with original Recurrent Convolutional Neural Network(RCNN), our method can reduce the loss of test model by 5.33%. And it's an effective method in improving the robustness of translation and rotation information than traditional VO methods.