Research On Scene Relocalization Algorithm Based On Cascaded Deep Neural Networks

Indoor localization plays an important role in people's life,work and study,and provides great convenience for people's clothing,food and shelter.The mainstream indoor localization technology uses different localization methods for different occasions,and faces various limitations such as power consumption,cost,and low localization accuracy in scene applications.With the continuous development of computer vision technology,deep learning methods provide a new vision-based relocalization technology for indoor scene relocalization.This paper proposes a scene relocalization algorithm based on cascaded deep neural network.Just input a single RGB picture,the camera pose of the picture can be given,and the scene relocalization can be realized.Through formal description of the indoor scene relocalization problem,a convolutional neural network is used to design a pose retrieval model,and a homoscedastic uncertainty loss is used for supervised training to allow the network to understand the various spatial information and semantics in the extracted scene.Information,etc.,to construct a pose retrieval feature database.Then,a point matching model based on deep neural network is proposed,so that the network can fully understand the point features in the extracted scene,instead of the traditional manual feature extraction and matching,and construct a map point feature database.Furthermore,using this cascade network,the pose features of the test pictures will be retrieved from the map pose feature database to the nearest Top K pictures,thereby greatly reducing the scene estimation range.Then the map point features corresponding to the retrieved library picture are matched with the test picture using the point feature extracted by the point matching network to provide a more accurate and efficient matching feature point.Finally,the multi-view geometry method is combined to estimate end-to-end.The camera pose of the picture realizes the relocalization of the indoor scene.The method of direct regression using the pose retrieval model will be used to perform the experiments under the same 7-Scenes data set as the algorithm using the cascade network for retrieval matching and combining multi-view geometry,and the existing representative algorithm.The results show that the method of searching and matching pose features using cascade neural network is better than the algorithm that directly uses neural network for regression.Compared with Pose Net and Pose LSTM,the translation error and rotation error are significantly reduced.The accuracy is improved significantly.