Visual Pose Estimation Based on the DenseNet Network

摘要

An end-to-end neural network model based on DenseNet was designed to estimate the pose of the camera in this paper. The picture frame captured by the camera and the camera position (3-dimensional space coordinates) and pose (quaternion) corresponding to the picture frame are the inputs to the network model. Through the neural network model, the spatial structure information and the higher-layer features in the image are trained and learned, so that the network model finally outputs the 7-dimensional vector representing the camera position (3-dimensional space coordinates) and the pose (quaternion). Due to the pose estimation constraint of the network, the training effect of the model is guaranteed, and the pose estimation ability of the network is improved. The trained model is validated on the StMarysChurch Dataset. The experimental results show that the network model has good performances in accuracy and shorter training time.