Research On Monocular Visual Odometry Based On Machine Learning

Autonomous navigation and localization technology is truly one of the key and difficult point for intelligent robot, and visual odometry technology has been one of scholars dedicated research direction. Compared with the traditional methods of SLAM, visual odometry is more concise, lower to the requirement of hardware; And compared with the wheel odometry, visual odometry is less affected by sideslip and kidnapping problem. However, visual odometry also suffers defects which affect its application, this paper will focus on these defects.In this paper, we study the basic design principle of monocular vis ual odometry, this paper proposes a feature point detection and matching algorithm framework, and it is compared with the SURF algorithm. In addition, this article discussed the methods to improve accuracy and stability of odomet ry from three aspects. First, in view of the monocular vision inherently suffers from scale ambiguity, this paper first puts forward a kind of Edge Expansion Method(EEM) based on watershed algorithm to automatically separate the ground from video. Secondly, in order to improve the the adaptive ability of noise for kalman filter in actual situations, this paper proposes a modified adaptive kalman filter algorithm based on machine learning algorithm. Third, considering the navigation and localization technology based on vision is usually affected by moving objects in the field of vision, this paper proposes that using the SVM + HOG to detect pedestrians, and eliminate the pedestrian feature points, so as to achieve the goal of anti-jamming.In this paper, the design of monocular visual odometry only need an ordinary camera as a sensor. It is simple assembly and cost low. Due to overcome the problems such as scale ambiguity, motion interference, we improve the ability of filter to resist noise, thus improve the accuracy and stability of the monocular visual odometry. It provides a new method for the design of visual odomet ry.