Multi-Granularity Environment Perception Algorithm Based On Optical Images

The environmental perception method is different from the offline traditional three-dimensional reconstruction.It emphasizes on using hardware platform which has low-cost and easy-access properties to understand the real working environment in real-time or near real-time,and to establish the environment model with needed precision according to the actual application needs.This concept so far has made great progress,and it is an important cornerstone of many popular areas,such as independent robot,three-dimensional reconstruction and augmented reality.At present,the most popular theoretical framework of environmental perception is simultaneous calculating the three-dimensional pose of camera and obtaining the three-dimensional structure information of the environment,named simultaneous localization and mapping.Because of the rich information contained in the optical image and the low price of the optical sensing device,the RGB image and IR image Based Simultaneous localization and mapping gets wide attention of many researchers.Although many works in this field have been made by many researchers during the last 30 years,there are still many robustness problems need to be solved urgently due to the limitation of computer performance,sensor noise,uncontrolled outdoor environment and diversified demands of practical application.To solve the above problems,this thesis adopts multi granularity probabilistic octree model and filtering theory to study the problem that how to achieve multi-sourced heterogeneous spatial optical image based fusion,multi-scale representation and single source error elimination method,and to explore the environmental perception problem especially in outdoor uncontrolled environment.The main works include:(1)Propose an optical image based multi granularity perception algorithm.The algorithm compresses and unifies the representation of multi granularity point clouds which fit the real 3D environment generated by kinds of optical images by using probabilistic octree model.At the same time,the algorithm utilizes the pruning and merging strategy to fit the multi granularity fusion and multi granularity representation requirements of environment modeling,meanwhile compresses the storage space of environment model.Then,the algorithm dynamically fuses the multi granularity point clouds of environment during the camera trajectory through the Kalman filter,and finally generates a unique temporal fusion probabilistic octree model,referred as TFPOM.The TFPOM is less likely influenced by noise and with arbitrary granularity to dynamically fit the real environment.Finally,the algorithm is evaluated by the performance of visual navigation environment model and augmented reality.The RMSE and the quality recovered from motion blur are used to quantitatively measure the performance of the proposed algorithm in the uncontrolled environment.(2)Propose a perception enhancement algorithm with filtering for the outdoor augmented reality scene.TFPOM based environment perception quality in outdoor uncontrollable scenarios is vulnerable to the effects of light and environment texture condition.Therefore,algorithm filters the image pyramids of video frames by using Laplacian of Gaussian operator to enhance the edges and details in the scene.Then in the internal frames matching process,algorithm utilizes the value of average of final accept score and average search length to ensure the similarity threshold in order to fit the change of contrast ratio between filtered and unfiltered video frames.Then the algorithm further improves the matching accuracy by using bilateral optical flow method.Finally,a snapshot of thumbnail based rapid relocation method is used by the proposed algorithm to restore the pose estimation of visual simultaneous localization and mapping when it faces tracking lost status.