Research and design of mobile robot slam system based on binocular vision

Mobile robot slam system based on binocular vision-Jake.cai This is a recent graduate student to do a reply to do the collation, PPT and related thinking map I have uploaded, the topic and direction is also my own thinking, in mobile robot slam technology I am completely white, reference to some domestic and foreign literature, The patchwork and the open topic to deal with the past, then said back, this slam is really very interesting. Don't say anything else, write it down for later viewing. ppt:http://download.csdn.net/download/jake_cai/9738397 Guide Map: http://download.csdn.net/detail/jake_cai/9738404
1. Overview

First, introduce slam, in our homes, offices, or factories and other indoor environment, mobile robots to really adapt to the unstructured environment, must have three elements of their own position, posture, environmental map, and in the general unfamiliar environment, these three elements are not available, Our slam is the key technology to solve the simultaneous localization and map construction of mobile robot. This paper mainly analyzes the problems of mobile robot in the indoor environment and map construction, and puts forward its own solution. 2. Ask the question-existing mobile robot simultaneous localization and map construction problems:

· The existing vision based mobile robot slam system for complex graphics processing, so the hardware configuration requirements of the robot is very high, need to match the GPU and other graphics processors. Increased the cost of the design of the robot. (complexity and cost of hardware)

· The existing mature Monocular Vision mobile Robot slam system cannot obtain the depth information of the environment directly through the image. But after the form of the algorithm to reply to the scene depth information and there is a problem of low precision. (For Monocular Vision Mobile robot)

· The existing mobile robot slam algorithm needs to recover the motion information from the image information, it is difficult to realize the accurate motion estimation, and the computation quantity is very great. (use visual odometer's disadvantage)

· The existing feature extraction algorithm based on SIFT (scale invariant feature extraction) generates 128-D feature descriptors, which will require a large amount of computation in stereo matching and data correlation stage.     (Actual application scenario) 3.        For the above problems, put forward solutions. The main solution: Hardware design complexity + SLAM algorithm real-time problem.

· The binocular camera built by the wireless video domain network in the laboratory is used to obtain the environment video and upload it to the PC, using the calculator to do the image processing and the pose estimation and the map construction and storage in the slam algorithm. It greatly reduces the hardware design difficulty and cost of the robot platform.

· Using the parallel calibrated binocular camera, the scene depth information is calculated directly by triangulation principle.

· The encoder and electronic compass are introduced into the equation of motion state, and the position and attitude are estimated. Avoid the huge amount of computational problems associated with the use of visual odometer.

· Improved SIFT algorithm, according to the distribution range of feature points, intercept the area of interest, reduce the number of redundant images to bring the calculation. At the same time, combined with concrete practical application, in the calibrated binocular vision camera, the rotation and scale transformation of the left and right images is small, so it can reduce the layer of image pyramid in the image scale space of sift algorithm.

4. System composition

Mainly using the research results of senior laboratory, based on video domain network to build a real-time transmission environment video to the PC side of the binocular camera, the camera node will collect video real-time transmission to the PC, and then use the computer as our main robot processor. The design complexity and cost of the robot are reduced.

Using a removable car with a encoder, an electronic compass sensor, a simple mobile robot platform can be built. 5. The whole frame of binocular vision robot Slam system

This is mainly divided into 4 pieces, data acquisition and transmission layer, image processing layer, positioning and map characteristics of the construction, executing agencies, now do the work is mainly hardware and image processing level of things. The algorithm of position and pose estimation based on Kalman filter is also studied. The construction of feature map is to build a sparse map based on feature point which is convenient for computer to compute and store. 6. The work flow of binocular vision mobile robot slam system

Here is the workflow for the entire slam system, which is divided into two parts, the key steps are:

· Based on the SIFT feature extraction algorithm, aiming at the practical application, we first intercept the image area of interest, then reduce the layer of image pyramid in the image scale space, and reduce the dimension of the feature description vector. To reduce the amount of calculation.

· In the process of stereo matching, the whole matching process is transformed into two-fork search problem in K-dimensional space using Kd-tree data organization. Improve the efficiency of matching.

· In the position and pose determination part, the extended Kalman filtering algorithm is used to fuse the information collected by the visual information and the internal encoder and the electronic compass. The minimum variance estimation of the current position and attitude of our mobile robot is obtained.

· The next step is the data association, which is the process of associating the landmark information extracted at different time position sensors, also known as the heavy-viewing process. In this study, we used the Euclidean distance method to do feature correlation. 7. Camera Calibration

The purpose of camera calibration is to obtain the internal parameter matrix and the external parameter matrix of the camera, and with these two matrices, we can achieve the mapping relationship between the world coordinate system and the camera coordinate system to the image coordinate system. In turn, the coordinates of the objects in the world coordinate system can be deduced from the image points in the picture. 8. Feature Extraction

· This is mainly to improve the SIFT algorithm, so that our feature points with rotation, scaling, light and other invariance.

· Image pyramid construction process, using different kernel size Gaussian filter for the original image convolution operation, and then the downward sampling. To obtain different scales, different sizes of graphics, forming an image pyramid.

· The vector description of the feature point is to establish a 4*4*8=128 dimension vector descriptor for the feature points extracted from the front. 9. Stereo Matching

Here we use a stereo matching algorithm for local area, using the image feature points extracted from the front, according to the imaging principle of the camera, looking for the same object point in the left and right two images, in order to reduce the matching operation quantity, the Polar line constraint condition is used, that is, the pair of conjugate pixels in the left and right two images must be above each other This can turn a two-dimensional plane search into a one-dimensional plane search problem. 10. Depth information calculation

This paper mainly uses the result of the last step stereo matching, according to the triangulation principle, calculates the parallax of a conjugate point on the left and right two images to determine the depth of the point. One-slam system model analysis

Now there are two kinds of analysis models for the slam problem, one is to deal with the problem of state estimation, the other is to deal with the optimization problem based on graph theory. Here I use the extended Kalman filtering algorithm to estimate the motion state of the system using a state estimation model with less computational overhead.

In the state equation, there are two components, the displacement of the robot relative to the initial point, and the rotation relative to the positive direction of the X axis. W is the process noise, which satisfies the Gaussian distribution.

The component of the vector z in the observation equation is the displacement and the rotation angle measured by the encoder and the electronic compass. and the displacement and rotation angle of the binocular vision system based on the feature point signpost angle between the frames.

The ultimate purpose of the SLAM process is to update the position estimation information of the robot. Because the robot's position and posture information can be estimated by the robot motion model, the robot's position can be corrected by using the robot's motion model to get the pose estimation, and the surrounding environment signpost information obtained by binocular vision system should be combined.