Design And Implementation Of Parallel Visual Odometer Based On Line Tracking

In recent years,with the rapid development of science and technology,artificial intelligence,driverless and augmented reality technologies have become a reality,and some of them even become an indispensable part of people's lives.The application of these techniques requires sensors to estimate their position in space and the threedimensional structure of the surrounding environment in real time.Considering factors such as price and scalability,video sequences based localization and environment awareness systems are widely used.In the Visual Simultaneous Localization and Mapping(V-SLAM)system,the Visual Odometry(VO)design for camera positioning is very important.The visual odometer consists of two parts: the front end and the back end.The front end is mainly responsible for tracking and mapping.The back end is responsible for eliminating the accumulated error caused by the front end and calculating the optimal local map.However,At present,the classic visual odometer front end and back end have some defects.The disadvantage of feature-based visual odometer front-end is that it is hard to construct semi-dense or even dense maps in real time;The lack of semantic information contained in feature points will cause mismatch problems and consumes a lot of computing resources.It is also redundancy in the optimization calculation of the map points at the back end of the visual odometer,and when the V-SLAM system is used across platforms,the use of the optimized library also faces many inconveniences such as the configuration environment and the installation of related libraries.Based on the classic algorithm of visual odometer,this paper proposes a visual odometer design scheme based on parallelized line tracking for the problems of front end and back end.By studying the influence of sampling point distribution on the accuracy of direct visual odometer,the characteristic points in the line are designed as sampling points,the feather point method is used to calculate the camera's rough position,which is used as the initial value for parallel line tracking.The solution not only can achieve stable camera positioning and solve the problems faced by the visual odometer back-end cross-platform,but also make full use of the semantic information of the lines in the image and the multi-threading advantages of multi-core computers,thus accelerating the localization of the visual odometer.Specifically,the main contributions of the paper are as follows:(1)For the front-end,this paper proposes a monocular parallelized line tracking algorithm based on LK optical flow method.Under the condition of small camera motion speed,the algorithm can achieve accurate matching of lines.Due to the limitations of the optical flow method and the change of the line shape caused by object occlusion,the camera cannot be successfully tracked when the camera moves at a high speed.(2)Proposes an improved algorithm for the stereo and RGBD camera.The stereo solution can build semi-dense depth maps in real time,and the RGBD solution can achieve more accurate matching of lines and obtain the pose of the camera at the same time.(3)For the back-end,this paper proposes a seven-step parallel algorithm and the derivation process of this algorithm is given.This algorithm can not only avoid the use of general-purpose optimization library to realize a more convenient cross-platform solution,but also adopt parallel computing method to speed up the optimization.(4)Design and implement a thread pool based parallel direct-feature hybrid RGBD visual odometer,which has better stability than the feature-based method and better accuracy than the direct method,and the speed of the algorithm is obviously improved when the thread pool is opened.