VINS（二）Feature Detection and Tracking

系统入口是feature_tracker_node.cpp文件中的main函数

1. 首先创建feature_tracker节点，从配置文件中读取信息（parameters.cpp），包括：

• ROS中发布订阅的话题名称；
• 图像尺寸；
• 特征跟踪参数；
• 是否需要加上鱼眼mask来去除边缘噪点；

%YAML:1.0

#common parameters
imu_topic: "/imu0"
image_topic: "/cam0/image_raw"

#camera calibration
model_type: PINHOLE
camera_name: camera
image_width:
image_height:
distortion_parameters:
   k1: -2.917e-01
   k2: 8.228e-02
   p1: 5.333e-05
   p2: -1.578e-04
projection_parameters:
   fx: 4.616e+02
   fy: 4.603e+02
   cx: 3.630e+02
   cy: 2.481e+02

# Extrinsic parameter between IMU and Camera.
estimate_extrinsic:    #   Have an accurate extrinsic parameters. We will trust the following imu^R_cam, imu^T_cam, don't change it.
                        #   Have an initial guess about extrinsic parameters. We will optimize around your initial guess.
                        #   Don't know anything about extrinsic parameters. You don't need to give R,T. We will try to calibrate it. Do some rotation movement at beginning.
ex_calib_result_path: "/config/euroc/ex_calib_result.yaml"  # If you choose  or , the extrinsic calibration result will be written vins_folder_path + ex_calib_result_path.
#If you choose  or , you should write down the following matrix.
#Rotation from camera frame to imu frame, imu^R_cam
extrinsicRotation: !!opencv-matrix
   rows:
   cols:
   dt: d
   data: [, -, ,
           , , ,
           , , ]
#Translation from camera frame to imu frame, imu^T_cam
extrinsicTranslation: !!opencv-matrix
   rows:
   cols:
   dt: d
   data: [-0.02,-0.06, 0.01]

#feature traker paprameters
max_cnt:             # max feature number in feature tracking
min_dist:             # min distance between two features
freq:                 # frequence (Hz) of publish tracking result. At least 10Hz for good estimation. If set , the frequence will be same as raw image
F_threshold: 1.0        # ransac threshold (pixel)
show_track:            # publish tracking image as topic
equalize:              # if image is too dark or light, trun on equalize to find enough features
fisheye:               # if using fisheye, trun on it. A circle mask will be loaded to remove edge noisy points

#optimization parameters
max_solver_time: 0.04  # max solver itration time (ms), to guarantee real time
max_num_iterations:    # max solver itrations, to guarantee real time
keyframe_parallax: 10.0 # keyframe selection threshold (pixel)

#imu parameters       The more accurate parameters you provide, the better performance
acc_n: 0.2          # accelerometer measurement noise standard deviation. #0.2
gyr_n: 0.02         # gyroscope measurement noise standard deviation.     #0.05
acc_w: 0.0002         # accelerometer bias random work noise standard deviation.  #0.02
gyr_w: 2.0e-5       # gyroscope bias random work noise standard deviation.     #4.0e-5
g_norm: 9.81007     # gravity magnitude

#loop closure parameters
loop_closure:    #if you want to use loop closure to minimize the drift, set loop_closure true and give your brief pattern file path and vocabulary file path accordingly;
                     #also give the camera calibration file same as feature_tracker node
pattern_file: "/support_files/brief_pattern.yml"
voc_file: "/support_files/brief_k10L6.bin"
min_loop_num: 

该config.yaml文件中的其他参数在vins_estimator_node中被读取，属于融合算法的参数。

• 优化参数（最大求解时间以保证实时性，不卡顿；最大迭代次数，避免冗余计算；视差阈值，用于选取sliding window中的关键帧）；
• imu参数，包括加速度计陀螺仪的测量噪声标准差、零偏随机游走噪声标准差，重力值（imu放火星上需要改变）；
• imu和camera之间的外参R，t；可选（0）已知精确的外参，运行中无需改变,（1）已知外参初值，运行中优化，（2）什么都不知道，在线初始化中标定
• 闭环参数，包括brief描述子的pattern文件（前端视觉使用光流跟踪，不需要计算描述子），针对场景训练好的DBow二进制字典文件；

2. 监听IMAGE_TOPIC, 有图像信息发布到IMAGE_TOPIC上时，执行回调函数：

ros::Subscriber sub_img = n.subscribe(IMAGE_TOPIC, , img_callback);

3. img_callback()

前端视觉的算法基本在这个回调函数中，步骤为：

　　1. 频率控制，保证每秒钟处理的image不多于FREQ；

　　2. 对于单目：

　　　　1）. readImage;

　　　　2）. showUndistortion(可选);

　　　　3）. 将特征点矫正（相机模型camodocal）后归一化平面的3D点（此时没有尺度信息，3D点p.z=1），像素2D点，以及特征的id，封装成ros的sensor_msgs::PointCloud消息类型;

　　3. 将处理完的图像信息用PointCloud和Image的消息类型，发布到"feature"和"feature_img"的topic：

pub_img = n.advertise<sensor_msgs::PointCloud>("feature", );
pub_match = n.advertise<sensor_msgs::Image>("feature_img",);

4. 包含的视觉算法：

1. CLAHE(Contrast Limited Adaptive Histogram Equalization)

cv::Ptr<cv::CLAHE> clahe = cv::createCLAHE(3.0, cv::Size(, ));

2. Optical Flow(光流追踪)

cv::calcOpticalFlowPyrLK(cur_img, forw_img, cur_pts, forw_pts, status, err, cv::Size(, ), );

3. 根据匹配点计算Fundamental Matrix， 然后用Ransac剔除不符合Fundamental Matrix的外点

cv::findFundamentalMat(un_prev_pts, un_forw_pts, cv::FM_RANSAC, F_THRESHOLD, 0.99, status);

4. 特征点检测：goodFeaturesToTrack， 使用Shi-Tomasi的改进版Harris corner

cv::goodFeaturesToTrack(forw_img, n_pts, MAX_CNT - forw_pts.size(), 0.1, MIN_DIST, mask);

特征点之间保证了最小距离30个像素，跟踪成功的特征点需要经过rotation-compensated旋转补偿的视差计算，视差在30个像素以上的特征点才会去参与三角化和后续的优化，保证了所有的特征点质量都是比较高的，同时降低了计算量。